The research was presented in Frontiers in Robotics and Artificial Intelligence in 2026 and evaluated both an artificial-intelligence perception system and its integration with a physical arm. The machine is designed for household-assistance applications, in which a user might ask a robot to retrieve one particular item from a table surrounded by clutter. Its arm weighs 2.2 kg, has a payload capacity of 1.5 kg and ends in a one-degree-of-freedom gripper that closes its thumb and four remaining fingers. The arm comprises three coupled tendon-driven modules: a three-degree-of-freedom shoulder, a two-degree-of-freedom elbow and a two-degree-of-freedom wrist. Its lightweight, elastic tendons were chosen partly to suit operation close to people.

To perceive the scene, the researchers converted an RGB-D image into a three-dimensional point cloud and used only each point’s spatial coordinates, excluding color information to reduce sensitivity to texture and lighting. A shared backbone extracted features for two linked tasks: estimating full six-degree-of-freedom grasp poses and assigning points to semantic object categories. The latter task links each possible grasp to a recognized object, allowing a user to select a target category in a graphical interface. The system then executes the highest-scoring grasp associated with that category, defined by the candidate point, approach direction, rotation around the approach axis, gripper width and gripper depth.

The central network innovation was a 3D Mixture-of-Experts UNet based on the MinkUNet architecture. Rather than activating an entire enlarged model, the design used eight parallel expert networks and a gating mechanism that selected the top two for each input. The experts replaced convolutional layers in the two deepest residual blocks of the encoder, where features represent broad object geometry and scene context. An auxiliary load-balancing loss discouraged the gate from routing nearly all inputs to a small subset of experts. Evaluation showed that the eight experts were selected with broadly balanced frequency, indicating that expert collapse did not occur, although tests of alternative expert counts, routing arrangements and block placements found no configuration that consistently dominated the chosen design.

The multi-task model contained 24.2 million parameters in total, of which 17.1 million were active during inference because only two experts ran at once. By comparison, a separate semantic-segmentation model using the study’s conventional 3D UNet backbone would have required at least 14.7 million additional active parameters. The model was trained for 40 epochs with a batch size of four on a single NVIDIA GeForce RTX 4080 SUPER graphics processor. It used cross-entropy losses for objectness and semantic classification, smooth L1 loss for continuous grasp quantities, and several task-specific terms for approach direction, gripper dimensions and grasp quality. Label smoothing was applied to segmentation with a coefficient of 0.1.

Benchmark testing used GraspNet-1Billion, a dataset of 97,280 RGB-D images from 190 cluttered real-world scenes containing 88 general objects and more than 1.1 billion grasp-pose annotations. The 190 scenes were divided into 100 training scenes and 90 test scenes, with the test scenes further separated into seen, similar and novel object groups. Because many of the objects had similar shapes, the researchers assigned the 88 objects to seven geometric categories: pump bottle, tube, container, irregular-shaped object, spherical object, elongated or cylindrical object, and cuboid object. Camera-specific train and test sets were evaluated separately for RealSense and Kinect data. The proposed backbone generally improved grasp estimation and segmentation over 3D UNet and PointNet++ alternatives, with gains reaching 5.24 average-precision points for grasp estimation and 3.42 mean intersection-over-union points for segmentation.

Those aggregate improvements did not apply uniformly. Compared with the 3D UNet backbone, the Mixture-of-Experts model improved RealSense grasp-pose average precision by 2.43 points on seen objects, 1.98 on similar objects and 2.14 on novel objects. For Kinect data, it improved seen and similar results by 5.24 and 2.25 points, respectively, but the novel-object result fell by 0.65 points. Segmentation similarly declined by 1.38 mIoU points on the RealSense novel subset, even as grasp-pose performance improved on that subset, suggesting a trade-off between the tasks’ shared representations. The authors associated weaker Kinect performance with sensor characteristics such as depth noise, missing points and differences in point-cloud density. On the similar and novel test sets, segmentation results were affected by systematic inter-class confusion: cuboid objects could be classified as containers, while pump bottles could be confused with tubes.

The arm itself used a geometry-based inverse-kinematics method to resolve its redundant seven-degree-of-freedom structure. When the end-effector pose is fixed, the elbow can occupy multiple positions, so the method introduced an arm angle and searched its permitted range in fixed increments. It selected the angle that minimized the sum of the absolute joint angles, producing a complete set of joint commands. The motion planner then interpolated position linearly and orientation with spherical linear interpolation. In the physical workflow, the end-effector first moved to a pre-grasp pose offset opposite the estimated approach vector, advanced to the selected grasp, lifted the object vertically, transferred it to a predefined placement location and released it.

The laboratory evaluation involved 10 target objects in scenes containing at least four objects each, with object numbers, positions and orientations varied manually between trials. Each target received 10 pick-and-place attempts, and success required grasping the specified object without toppling it, lifting and transporting it, then releasing it without a drop. All three seen objects achieved 100% success. Among unseen objects, the individual rates ranged from 90% to zero: objects 4 and 6 each succeeded in nine trials, objects 7 and 9 in seven, object 5 in four, object 8 in one, and object 10 in none. The 52.9% average for unseen objects therefore concealed substantial differences associated with object geometry, size and material. The researchers identified several contributing constraints, including incomplete depth measurements from translucent plastic, limited contact options imposed by the one-degree-of-freedom gripper, and difficulty avoiding the supporting surface when few stable grasp regions existed.

The complete system, including the robotic-arm controller, used 5.1 GB of RAM and 3.12 GB of video memory and processed images at an average of 20 FPS in the experimental setup. These measurements suggest that the workload could fit on a compact edge computer such as a Jetson Orin2, but its performance on such hardware was not tested. The prototype also had important functional limits. Semantic categories allowed a user to request a broad object type, yet two instances of the same type could not be selected separately without instance-level segmentation. The study proposes future work on more adaptable multi-degree-of-freedom hands, improved depth acquisition for translucent objects, better mixture-of-experts tuning and instance segmentation. It also suggests voice or electromyography-based target selection for users with limited physical mobility, an application direction that was proposed rather than evaluated.

Subject of Research: A 7-DOF robotic system for autonomous perception and grasping of user-selected general objects

Article Title: Development of a 7-DOF robotic arm control system for autonomous grasping of user-specified general objects

Article References: Bai, S., Jiang, Y., Yokoi, H., & Togo, S. (2026). Development of a 7-DOF robotic arm control system for autonomous grasping of user-specified general objects. Frontiers in Robotics and AI, 13, 1955232. https://doi.org/10.3389/frobt.2026.1955232

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1955232

Keywords: robotic grasping, assistive robots, 7-DOF arm, point-cloud perception, grasp pose estimation, semantic segmentation, mixture of experts, inverse kinematics, tendon-driven robot, pick-and-place, Development, robotic

The study targeted a difficult form of image-based plant phenotyping: identifying thin, branching structures when the foreground occupies only 2.43% of all annotated pixels. Tobacco leaves are useful test subjects because their broad lamina contains a prominent midrib and dense lateral network, while curing continually changes their color, moisture, texture, area, and vein visibility. Images collected from May through September 2025 covered 28 curing sequences, six cultivars, four producing regions in China, both upper and lower leaf surfaces, and multiple stages from green leaves through yellowing, browning, and drying. A custom imaging system captured 2,560 × 1,920-pixel photographs every 10 minutes inside curing barns.

For model development, the researchers reduced the sequence sampling because leaf appearance changed little during each 10-minute interval. They reserved 50 images from one independently selected adaxial group and 49 from one abaxial group for testing. The remaining 891 images were divided into 792 training images and 99 validation images. Horizontal flipping, random rotation, brightness adjustment, and Gaussian blur expanded the training set to 3,960 samples, with masks transformed alongside images for the geometric changes. Vein masks were manually drawn under a predefined protocol that excluded ambiguous textures and wrinkles unless their continuity with visible venation could be confirmed. Experienced reviewers checked the annotations, which were converted to binary images for training. Quantitative evaluation was conducted on network outputs resized to 256 × 256 pixels without preserving the original 1,060 × 1,920 aspect ratio.

CTVSeg-UNet builds on U-Net, an encoder-decoder network whose skip connections carry fine spatial information from the image-analysis stage to the reconstruction stage. Three modifications address the scarcity of vein pixels and the need to retain narrow branches. First, the cross-entropy loss assigned veins a weight of 40 and background a weight of 1, increasing the influence of the minority class during optimization. Second, a collaborative upsampling module ran bilinear interpolation and transposed convolution in parallel and fused their outputs. Bilinear interpolation offered stable spatial recovery, while learnable transposed convolution could reconstruct finer details. Third, the enhanced U-Net served as the generator in a PatchGAN adversarial framework, with a discriminator evaluating local consistency between leaf images paired with real or predicted masks.

Comparisons with four semantic-segmentation models showed why conventional accuracy figures alone could be misleading. PSPNet achieved 60.33% mean pixel accuracy and 58.64% mean intersection over union, while DeepLabV3 reached 60.46% and 58.83%. SegNet scored 64.37% and 60.34%. The baseline U-Net rose to 73.26% mean pixel accuracy and 67.00% mean intersection over union, but its visual masks contained broken branches and missed fine laterals. CTVSeg-UNet produced the most visually complete networks, recovering most of the midrib and many lateral veins across stages and cultivars. Its mean intersection over union was 65.60% and its F1 score 50.73%, both slightly below the baseline U-Net’s 67.00% and 52.48%.

The model’s lower precision, 37.82%, reflected its greater willingness to identify faint vein-like structures. DeepLabV3 had the highest precision at 69.36%, yet it detected only portions of the midrib and missed most lateral veins. CTVSeg-UNet instead classified more adjacent lamina and boundary pixels as veins. The authors also note that some extremely faint branches may have been omitted from the manual masks, meaning a portion of the apparent false-positive area could correspond to genuine venation. A miss-rate analysis found that 13.39% of annotated vein pixels went unidentified by CTVSeg-UNet, compared with 50.15% for the baseline U-Net. The refined model also produced lateral-vein spacing estimates closer to the ground truth, including for fine branches that the baseline could not measure.

Ablation experiments demonstrated that the improvements worked mainly as a combined system. Replacing transposed convolution with bilinear interpolation alone barely changed baseline performance, whereas the collaborative module raised mean intersection over union to 68.03%, precision to 63.08%, and F1 score to 54.58%. Adding the 40:1 class weights increased recall from 47.12% to 64.60% and mean pixel accuracy from 73.26% to 81.63%, but reduced precision to 47.33%. Adversarial learning applied directly to the unmodified U-Net lowered recall to 42.14%, showing that a discriminator could not compensate for a generator with poor vein-recognition ability. Combining collaborative upsampling and weighted loss raised recall to 72.46%; adding adversarial refinement then increased it to 77.03%. Three repeated runs of the final and penultimate configurations gave similar results, indicating stable training in those comparisons.

The researchers translated the model into an application that links camera acquisition, image separation, standardization, saturation-channel extraction, inference, mask display, and local storage. In batch mode, it processes previously collected curing images. In real-time mode, an edge device connected through TCP/IP controls the camera, preprocesses each frame, runs the embedded model, displays the original image and predicted mask, and stores the results during curing. Such masks can feed subsequent measurements of vein continuity, lateral spacing, and the angle between the midrib and lateral veins. The deployment demonstrates workflow feasibility, but the study did not report a formal field trial of curing decisions or end-to-end improvements in tobacco quality.

Visual tests across all six cultivars and both leaf surfaces showed that the system generally preserved the dominant venation framework despite differences in morphology, texture, and viewing angle. Performance nevertheless declined during the late color-fixing and stem-drying stages, when dehydration produced wrinkles and folds and reduced contrast between fine veins and the surrounding lamina. Some masks became locally discontinuous, missed lateral branches, or treated folded textures as veins. These limitations identify the main targets for future work: temporal information from curing sequences, multi-view imaging, structure-aware evaluation, and topology-preserving constraints. The study involved plant images only, with no human participants or live-animal experiments. Overall, the results show that CTVSeg-UNet can extract more complete tobacco vein networks than the tested baseline, while its lower precision and reduced performance in late-stage leaves remain important limitations of the reported evaluation.

Subject of Research: Deep-learning segmentation of tobacco leaf vein networks during flue curing

Article Title: CTVSeg-UNet: robust refinement for leaf vein phenotyping in curing tobacco

Article References: Ma, Y., Zhang, H., Wang, W., Zhang, X., Zhang, M., Zhang, H., Song, Z., & Wu, J. (2026). CTVSeg-UNet: robust refinement for leaf vein phenotyping in curing tobacco. Smart Agricultural Technology, 15, Article 102650. https://doi.org/10.1016/j.atech.2026.102650

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102650

Keywords: CTVSeg-UNet, tobacco curing, leaf venation, semantic segmentation, plant phenotyping, U-Net, computer vision, adversarial learning, thin structures, robust, refinement, leaf

The experiment involved 100 Yellow and Spotted-brown chickens at Fuhe Chicken Farm in Guangzhou, Guangdong Province, China. The population included 70 Spotted-brown chickens aged 163 days, 15 Spotted-brown chickens aged 144 days and 15 Yellow chickens aged 183 days. Each bird underwent 10 non-consecutive captures, with every visible-light frame acquired synchronously with an infrared frame. The collection yielded 1,000 images per modality and 1,000 spatially registered, fused images. The researchers partitioned the birds, rather than their images, at an 8:1:1 ratio into training, validation and independent test sets, ensuring that all repeated views of a given chicken remained in the same subset. Two operators measured shank length manually with vernier calipers, and their average served as the individual reference used for correlation and agreement analyses.

The imaging platform combined a HIK MV-CE060–10UC visible-light industrial camera, operating at 3,072 by 2,048 pixels and 42.7 frames per second, with an IRay AT-61 uncooled vanadium-oxide infrared camera operating at 640 by 512 pixels and 25 frames per second. The optical distance from the camera center to the chicken’s suspended measurement plane was fixed at 380 mm. Two custom light arrays illuminated the scene, and a heated double-layer stainless-steel calibration plate with a 20 mm center-to-center spacing provided a stable geometric reference. Corresponding circular patterns in the calibration plate established control points for registering the infrared images to the visible-light coordinate system. A Haar-wavelet fusion procedure then combined visible high-frequency texture with infrared low-frequency thermal information. The resulting conversion factors were 0.1770 mm per pixel for visible and fused images and 0.3509 mm per pixel for the original infrared images.

The central technical change was the replacement of direct scalar regression with pixel-level anatomical segmentation. SegFormer-B1 used a Mix Transformer B1 encoder and a lightweight all-multilayer-perceptron decoder to classify each image pixel as background, an upper bare metatarsal-toe region or a lower feathered thigh-shank region. The models were trained for 6,400 iterations with AdamW on 640-by-640-pixel crops and a batch size of 24. Masked cross-entropy loss excluded invalid zero-padded pixels. Images were normalized, randomly resized over scales from 0.5 to 2.0, cropped and augmented with horizontal flipping and photometric distortion; the study reported their use but did not separately measure whether any particular augmentation improved performance. Because the segmentation masks and both endpoint locations remained available as intermediate outputs, an unusual measurement could be traced to semantic segmentation or upper- and lower-anchor localization rather than to an unexplained scalar output.

On the independent test set, fused images produced the strongest overall segmentation performance, with a mean intersection over union of 93.92%, overall accuracy of 99.54% and mean pixel accuracy of 96.62%. Visible images achieved corresponding values of 88.62%, 99.20% and 92.41%, while infrared images reached 92.79%, 99.45% and 95.61%. The largest difference among modalities occurred in the lower feathered region, where intersection over union was 87.61% for fused images, 85.36% for infrared and 75.51% for visible light. For the upper bare region, the values were 94.61%, 93.59% and 91.13%, respectively. The authors associated the fused result with complementary image information: visible texture helped constrain anatomical boundaries, while infrared contours helped compensate for contour loss caused by shadows and feather occlusion.

The two measurement anchors were localized with different rules because the two anatomical regions responded differently to inverted suspension and articulated movement. The upper key point was the global geometric centroid of the complete bare metatarsal-toe mask. Using all valid pixels in that region acted as a two-dimensional low-pass operation, reducing sensitivity to abrupt boundary changes caused by toes opening, closing or adopting different configurations. The lower key point was identified at the interface between the feathered and bare regions. A morphological difference operation extracted a single-pixel boundary from the feathered-region mask, and a distance transform measured each boundary pixel’s shortest geometric distance to the bare region. A two-pixel contact threshold isolated the interface set, whose vertical extreme became the lower anchor. At the calibrated plane, that threshold represented 0.3540 mm in the visible and fused modalities and 0.7018 mm in infrared images. The Euclidean distance between the two anchors was then converted from pixels to millimetres.

Ablation experiments indicated that both spatial constraints contributed to repeated-measurement stability. Replacing the global centroid with skeleton-junction localization increased the fused modality’s mean absolute deviation from 0.2894 to 1.2186 mm, its mean standard deviation from 0.3663 to 1.6641 mm and its mean range ratio from 1.54% to 6.45%. Its Pearson correlation with manual measurements also fell from 0.7109 to 0.6782. Visual comparisons showed that local changes in toe topology could shift a skeleton junction between captures, while the centroid remained near the center of the predicted upper region. Replacing the distance-transform interface with medial-axis peak localization produced a fused mean absolute deviation of 2.0929 mm, a mean standard deviation of 2.6454 mm and a mean range ratio of 12.81%. That alternative had a slightly higher Pearson correlation of 0.7458, but its anchor varied more across repeated images as feather coverage and regional geometry changed.

Visible images had the strongest association with the manual reference, with Pearson’s r of 0.7696 and p = 0.0092. Infrared images produced r = 0.7150 and p = 0.0201, while fused images produced r = 0.7109 and p = 0.0212. All three modalities therefore showed significant positive correlations at the individual level. The fused modality nevertheless delivered the most consistent measurements, with the lowest mean and median absolute deviations, standard deviations, coefficients of variation and range ratios. Its median absolute deviation was 0.2585 mm, its median standard deviation was 0.3354 mm, its median coefficient of variation was 0.44% and its median range ratio was 1.29%. The results support an increasing association between the visual and manual measurements and show that fusion reduced acquisition-to-acquisition fluctuation; they do not establish that the methods reproduced the complete ordering of individual birds’ shank lengths.

Repeated-measures Bland–Altman analysis established a clear boundary on the result. The fused system produced values lower than manual caliper measurements by a mean of 13.0411 mm, with 95% limits of agreement from 2.4178 to 23.6643 mm. The visible and infrared mean biases were 14.3377 and 20.0588 mm, respectively, with limits of agreement of 4.6586–24.0168 mm and 9.5217–30.5958 mm. Differing endpoint definitions were presented as a possible main explanation for the systematic difference: operators selected caliper contact points, sometimes after parting or compressing feathers, whereas the algorithm used the centroid of the bare upper region and the interface between two segmented regions. Regression found no significant proportional bias, and Spearman analysis found no significant heteroscedasticity within the tested range. The wide agreement intervals nevertheless showed that the amount of underestimation was not constant among chickens.

The study has important limitations. The independent test set contained only 10 birds and was unevenly distributed across two breeds and three ages. Ten repeated observations per chicken increased the amount of within-animal information but did not increase the number of statistically independent individuals, limiting the precision of the Bland–Altman estimates and the statistical power of the proportional-bias and heteroscedasticity tests. The design also could not separate the effects of breed, age and individual morphology on the wide limits of agreement. All birds were measured in a controlled suspended setup, leaving performance during free movement, under different handling conditions and in uncontrolled farming environments unknown. Larger, balanced and independently validated populations will be needed to establish whether the visual and caliper scales remain stable across broader chicken groups. Until then, the system is best regarded as a consistent contactless visual reference for phenotypic comparison, rather than an uncalibrated numerical replacement for manual caliper measurement.

Subject of Research: Contactless measurement of live-chicken shank length using multimodal semantic segmentation and geometric endpoint localization

Article Title: Multimodal semantic segmentation measurement of shank length in live chickens based on SegFormer and geometric positioning strategy

Article References: Zheng, H., Lin, Z., Ma, C., Lin, P., Liang, Z., Bi, W., Guo, Y., Jiang, T., Yuan, K., Ma, L., & Ding, H. (2026). Multimodal semantic segmentation measurement of shank length in live chickens based on SegFormer and geometric positioning strategy. Smart Agricultural Technology, 15, Article 102641. https://doi.org/10.1016/j.atech.2026.102641

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102641

Keywords: live chicken phenotyping, shank length, SegFormer-B1, semantic segmentation, infrared imaging, multimodal fusion, geometric localization, distance transform, computer vision, Multimodal, semantic, segmentation

The threat at the heart of the work is the membership inference attack, or MIA. In such an attack, an adversary interrogates a trained model and tries to determine whether a specific record, perhaps a patient’s medical scan or a person’s biometric template, was part of the data used to train it. The mere confirmation that someone’s data was in a hospital’s training set can be a privacy violation in its own right, particularly under regulations such as the GDPR that govern how personal information may be processed. Membership inference attacks were first formalized by Shokri and colleagues in 2017, and since then an entire research literature has grown around both sharpening these attacks and defending against them.

What has frustrated defenders is that most existing attack and evaluation methods rest on assumptions that rarely hold in the real world. Classic approaches require shadow models, replicas trained to mimic the target’s behavior, or access to confidence scores, or knowledge of how the data was split between training and test sets. In a genuine black-box scenario, where an evaluator can only query a deployed model and observe its outputs, these assumptions are often unrealistic. Moreover, existing techniques tend to be computationally expensive, and many were designed exclusively for classification tasks, leaving semantic segmentation models, which are central to medical imaging, poorly covered. Evaluating whether decentralized training paradigms such as federated learning resist these attacks has therefore remained a genuine open challenge.

The IIT Kharagpur team’s contribution is a framework that sidesteps these obstacles entirely. Their method assesses membership privacy risk through what they call dropout-induced inference deviation. Dropout, a technique originally introduced by Srivastava and colleagues in 2014 to prevent overfitting, randomly deactivates units in a neural network during training. The researchers exploit a different facet of dropout: by randomly dropping units at inference time and observing how much the model’s outputs fluctuate, they obtain a signal that distinguishes memorized training samples from unseen ones. Samples the model has memorized respond differently to this induced randomness than samples it has merely generalized over, and the magnitude of that deviation serves as a proxy for membership risk.

The elegance of the approach lies in its minimal requirements. It demands no shadow models, no auxiliary datasets, and no access to the model’s internal parameters or train-test splits. Anyone who can query a deployed black-box model can apply the diagnostic, making it a practical tool for organizations that need to audit privacy risk in systems they did not train. Because it works for both classification and segmentation tasks, it covers the full spectrum of vision models used in biometric authentication and medical imaging, the cross-disciplinary context in which the work is situated.

Beyond the empirical tool, the study offers a theoretical account of why federated learning should resist memorization in the first place. Through sensitivity analysis and an examination of loss surface geometry, the authors show that centralized training tends to converge to sharp minima, narrow valleys in the loss landscape where the model fits the training data with high precision. Sharp minima are associated with overfitting and with amplified output variability between training and non-training samples, which is precisely the variability an inference attacker exploits. Federated aggregation, in which multiple clients train locally and only model updates are combined by a central server, drives models toward flatter minima. Flatness, as prior work on generative models has also suggested, suppresses memorization and reduces the leakage of membership information.

This connection between loss landscape geometry and privacy is the conceptual payoff of the paper. It reframes federated learning not merely as a data-governance arrangement, in which raw data never leaves the client device or institution, but as a training procedure whose optimization dynamics themselves confer a privacy benefit. The aggregation of updates across heterogeneous clients acts as a form of implicit regularization, smoothing the loss surface and discouraging the model from latching onto individual records. That mechanistic insight elevates the work beyond a benchmarking exercise and gives practitioners a principled reason to prefer decentralized training when privacy is paramount.

The empirical validation is unusually broad. The framework was tested across seven datasets spanning both classification and segmentation tasks, including biomedical image benchmarks such as MedMNIST and the ISIC skin lesion archive, alongside classic computer vision datasets like CIFAR and SVHN. Model architectures were equally diverse, ranging from residual networks and MobileNet variants for classification to U-Net, U-Net++, and DeepLabV3+ for segmentation. Across this breadth, the dropout-induced deviation signal consistently separated training from non-training samples, and the comparison between centralized and federated training regimes confirmed that federated aggregation reduces membership leakage. Demonstrating generality across tasks and architectures is what distinguishes the method from earlier evaluation schemes that were tailored to a single model family or output type.

The practical implications extend well into the domains where the stakes are highest. In medical imaging, models trained on hospital data are exactly the kind of high-value, sensitive assets that membership inference attacks target, and regulations going back decades, from early federal medical privacy rules to modern data protection law, recognize the harm that even metadata-level disclosures can cause. In biometric authentication, where the training data consists of irrevocable personal identifiers, the ability to audit a deployed model for memorization without touching its internals is a significant operational advantage. A hospital or a device manufacturer could run the diagnostic on a production model, quantify its membership risk, and make informed decisions about whether additional defenses, such as input perturbation at the client level, are warranted.

The work also contributes to a broader recalibration in the security community. Recent systematizations of knowledge have argued that membership inference is harder than previously thought, and that many reported attack successes depend on favorable assumptions about data dependencies, distribution shifts, and attacker knowledge. Against that backdrop, a cheap, assumption-free measurement tool is valuable precisely because it lets defenders, not just attackers, quantify risk under realistic black-box conditions. The authors’ framework, computable with nothing more than repeated stochastic queries to the model, turns privacy auditing from an expensive research project into a routine check. As federated learning continues to spread through healthcare, mobile computing, and cross-institutional collaborations, the study suggests that the decentralized paradigm offers not just regulatory comfort but a measurable, geometrically grounded reduction in the risk that a model will give up the secrets of the data it was built on.

Subject of Research: Membership privacy risk assessment in federated versus centralized deep learning

Article Title: Can federated models keep secrets better?

Article References: Borthakur, A., Srivastava, A., & Sheet, D. (2026). Can federated models keep secrets better?. Multimedia Tools and Applications, 85(11), Article 816. https://doi.org/10.1007/s11042-026-21981-5

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21981-5

Keywords: federated learning, membership inference attack, deep learning, data privacy, dropout, loss surface geometry, black-box evaluation, medical imaging, biometric authentication, semantic segmentation, sharp minima, privacy auditing

The study, published in the Journal of Ovarian Research, comes from the Institute for Maternal and Child Health IRCCS Burlo Garofolo and the University of Trieste. First author Alessandro Biscontin and colleagues tackled a problem that has long frustrated embryologists: morphology-based assessment of metaphase II oocytes, the mature eggs used in intracytoplasmic sperm injection, is a cornerstone of ICSI workflows but suffers from inter-operator variability and limited reproducibility. Two experts looking at the same microscopic image can disagree on where one structure ends and another begins, and purely data-driven deep learning models inherit a different version of the same problem, generating anatomically implausible segmentations when the training data does not cover every visual contingency.

The researchers’ answer is a neurosymbolic architecture, a design philosophy that pairs the pattern-recognition power of neural networks with the explicit, human-readable constraints of symbolic artificial intelligence. In their system, a transformer-based segmentation model called SegFormer first predicts regions of interest in a microscopic image of a denuded MII oocyte. A logic score, derived from a set of knowledge-driven anatomical constraints, is then computed on those predicted regions. The constraints encode structural facts that embryologists know by heart: the zona pellucida forms a coherent ring around the cell, the perivitelline space lies between the oolemma and the zona pellucida, the first polar body sits within or adjacent to that space, and the cytoplasm occupies the interior. When the network’s predictions violate these relationships, the logic score drops.

What makes the approach technically elegant is how the logic score is used. During training, it functions as an auxiliary loss, computed alongside the standard segmentation loss, so that parameter updates are guided not only by pixel-level agreement with ground-truth annotations but also by whether the emerging segmentation is morphologically coherent. At inference time, the same score becomes a continuous measure of morphological reliability, a number that tells the user how anatomically plausible the model’s output is for a given image. The framework is model-agnostic in spirit: the authors also applied the logic loss to a convolutional backbone, showing that the benefit is not tied to the transformer architecture alone.

The dataset behind the study is substantial. The team retrospectively analyzed 1,243 microscopic images acquired at 200× magnification from 218 patients undergoing ICSI at a single reproductive medicine center. Written informed consent was obtained from all participants, and the study received ethical approval from the institutional review board under protocol IRB-BURLO 05/2022. Each image was annotated for the structures that matter in oocyte assessment: the zona pellucida, the perivitelline space, the first polar body, the cytoplasm, and the background, along with the holding and injection pipettes visible in the ICSI setup.

The quantitative results are striking. The neurosymbolic model achieved a macro-average intersection over union of 0.84 and a Dice similarity score of 0.91, outperforming an architecturally identical SegFormer baseline that reached an IoU of 0.80 and a Dice of 0.88. Classic convolutional architectures trained on the same dataset and split fared worse still: U-Net reached an IoU of 0.76, while DeepLabV3+ managed 0.81. In semantic segmentation, where IoU measures the overlap between predicted and true regions for each anatomical class, a four-point gain over an identical baseline is a meaningful improvement, and the gain came without any change to the network architecture or the amount of training data.

Perhaps the most important part of the study is the set of controls the authors designed to prove that the improvement comes from the content of the logical rules rather than from the mere presence of an auxiliary term. They trained two controls with identical loss form and weight but no anatomical meaning: a non-structural auxiliary loss and a scrambled rule set. The non-structural auxiliary loss of equal weight did not reproduce the gain, reaching an IoU of 0.81 and a Dice of 0.88, essentially matching the baseline. Scrambling the semantics of the rules degraded performance dramatically, dropping the IoU to 0.61. The message is unambiguous: the improvement is driven by what the constraints say, not by how they are mathematically packaged.

This control design addresses a subtle criticism that has dogged auxiliary-loss approaches in machine learning. It is easy to add a term to a loss function and observe a performance bump, but harder to demonstrate that the bump reflects genuine knowledge rather than a generic regularization effect. By showing that a meaningless auxiliary term of the same weight produces no benefit, and that scrambling the rules actively harms performance, the authors provide evidence that domain knowledge, when formalized as logical constraints, is a real and transferable ingredient of model training. The scrambled-rule result also hints at the framework’s sensitivity: constraints that contradict anatomy do not merely fail to help, they mislead the network.

The implications reach beyond oocyte imaging. Assisted reproductive technologies are increasingly supported by automated image analysis, from sperm selection to embryo grading, and the field faces the same twin challenges everywhere: deep learning models are powerful but opaque, and their predictions can fail in ways that clinicians cannot easily anticipate. A neurosymbolic layer offers a partial remedy. Because the constraints are explicit and human-readable, an embryologist can inspect them, verify that they match clinical understanding, and even modify them as knowledge evolves. Because the logic score is computed at inference time, it can serve as a continuous quality indicator, flagging predictions that are statistically plausible but anatomically suspect before they reach a decision-support workflow.

The authors are careful to position the framework as an image-analysis component for future decision-support systems in ICSI workflows rather than as a standalone clinical tool. The study was retrospective, single-center, and focused on denuded MII oocytes at a fixed magnification, so external validation across centers, magnifications, and imaging setups remains future work. Still, the demonstration that anatomical knowledge can be embedded as logical constraints and measurably improve both accuracy and morphological coherence marks a promising direction for artificial intelligence in reproductive medicine. As the authors conclude, embedding domain knowledge as logical constraints yields more accurate and morphologically coherent segmentations of MII oocyte images than a purely neural baseline, and that hybrid recipe, logic guiding learning, may prove just as valuable wherever medical image analysis meets clinical anatomy.

Subject of Research: Neurosymbolic deep learning with anatomical logical constraints for semantic segmentation of metaphase II human oocyte microscopy images

Article Title: Neurosymbolic approach improves robustness of deep learning models for semantic segmentation of MII human oocyte images

Article References: Biscontin, A., Ajčević, M., Luppi, S., Giolo, E., Miladinović, A., Bottega, R., De Rocco, D., Martinelli, M., Accardo, A., Skerk, K., Zito, G., & Ricci, G. (2026). Neurosymbolic approach improves robustness of deep learning models for semantic segmentation of MII human oocyte images. Journal of Ovarian Research. https://doi.org/10.1186/s13048-026-02286-5

Image Credits: AI Generated

DOI: 10.1186/s13048-026-02286-5

Keywords: neurosymbolic AI, deep learning, semantic segmentation, oocyte, ICSI, zona pellucida, transformer, medical image analysis, reproductive medicine, Journal of Ovarian Research, Neurosymbolic, approach

The research, led by Dian Liu of South China University of Technology together with collaborators at Salesforce, an independent research group in Palo Alto, and the University of Oulu, addresses a problem that has plagued earlier wrong-way detection systems: they assume the world stands still. Most existing camera-based approaches rely on fixed lane-direction settings that are calibrated once and never updated. But highway surveillance cameras are mounted on poles exposed to wind, vibration, and maintenance crews. A small physical rotation can silently invalidate the hard-coded lane regions and reference directions, causing ordinary vehicles to be misclassified as wrong-way targets. The consequences are not merely technical. False alarms erode operator trust in automated traffic safety systems, while missed detections defeat the entire purpose of early warning.

The new framework tackles this with two coupled modules: online dynamic calibration and cascaded filtering detection. The calibration module begins by using DeepLabv3+, a well-established semantic segmentation network, to extract drivable road areas and generate a binary road mask, which is then refined with morphological operations and clustered into individual lanes using connected-component labeling. Lane directions are not manually annotated. Instead, the system mines them from history: it aggregates displacement vectors from vehicles that have previously traversed each lane, compiling a standardized reference direction that reflects the statistically dominant flow. This design deliberately excludes the instantaneous motion of any current candidate vehicle, preventing a single anomalous trajectory from contaminating the reference itself.

The most striking innovation is how the system detects when its own viewpoint has been compromised. Rather than running computationally expensive optical flow or homography estimation on every frame, the framework employs difference hashing, or DHash, a technique that compresses each video frame into a 64-bit binary code by comparing the grayscale values of adjacent pixels in a downsampled image. By computing the average Hamming distance between hash codes of consecutive frames, the system can tell when a substantial perspective rotation has occurred. When the distance exceeds a configured threshold of 25.5, the framework concludes that the camera has physically moved, reinvokes the segmentation network on the post-rotation stream, reconstructs the lane partitioning, and updates the direction references accordingly. A five-frame dormancy lock prevents the mechanism from repeatedly firing during the unstable transition period of a single rotation event. The entire rotation-perception step costs less than 0.1 milliseconds per frame.

Once calibration is handled, the second module confronts a subtler enemy: pseudo-motion. Vehicles that are actually stationary can appear to move when their bounding boxes drift, when they are partially occluded by passing trucks, or when flashing hazard lights create local illumination changes that mimic displacement. The cascaded filtering pipeline addresses this in stages. First, a temporal continuity filter maintains a wrong-way status counter for every tracked vehicle. Only when the counter continuously grows across consecutive frames, exceeding a persistence threshold, does the target graduate from transient tracking noise into a suspected wrong-way candidate. Validation experiments showed this threshold matters enormously: with a threshold of one frame, precision was a dismal 68.29 percent, but at five consecutive frames it reached 96.77 percent with recall unchanged, striking the optimal balance between alarm latency and false-alarm suppression.

Suspected candidates then undergo a deeper semantic examination. The framework captures the start and end frames of each candidate trajectory and computes the deep feature difference between them using a VGG-16 network pretrained on ImageNet. The resulting feature difference map is normalized and cropped to the target’s bounding box, and the arithmetic mean of the feature differences within that region quantifies whether the vehicle actually moved during the interval. Stationary vehicles, the researchers found, produce feature difference values clustered around 0.21 to 0.25 on a normalized scale, while genuinely moving vehicles produce values from 0.75 to 0.97, a separation wide enough for a max-margin classifier to draw a robust decision boundary. The classifier is trained offline on a dedicated feature-difference dataset using a support-vector-machine-style optimization with a non-linear kernel, solved as a convex quadratic programming problem.

To evaluate the system under realistic conditions, the team constructed a dedicated highway surveillance dataset of 99 video sequences drawn from operational cameras, spanning open highways, tunnels, and deliberate camera-rotation events. The sequences were partitioned at the video level into 51 training, 20 validation, and 28 independent test videos, comprising more than five hours of footage and 43 independent wrong-way events. Frames from the same video never appeared in both tuning and testing sets, preventing data leakage. The videos were sampled down to approximately 2.08 frames per second to emulate the low-frame-rate streams typical of roadside edge deployments, and the feature-difference dataset contained 20,864 training samples of stationary and moving vehicles for the secondary verification module.

The results are striking. Across the full independent test set, the proposed framework achieved 97.62 percent precision, 95.35 percent recall, and a 96.47 percent F1-score, with an operational false alarm rate of just 0.12 false alarms per hour. The advantage was most dramatic under camera rotation, the scenario that cripples conventional systems. There, the framework maintained 93.33 percent recall and precision with only 0.15 false alarms per hour, while baseline methods relying on static direction references collapsed to F1-scores between 40 and 59 percent with false alarm rates as high as 4.2 per hour. Ablation studies confirmed that every component earns its place: removing the dynamic calibration module dropped camera-rotation recall from 93.33 to 46.67 percent, while removing temporal filtering or feature-difference verification each degraded overall performance substantially.

The system also runs fast enough for practical deployment. On a workstation equipped with an NVIDIA RTX 2080 SUPER GPU, the complete pipeline achieved an average throughput of 48.5 frames per second, or about 20.6 milliseconds per frame including video decoding and memory transfer overhead. Object detection via YOLOv8 consumed 9.8 milliseconds per frame, tracking via ByteTrack 2.6 milliseconds, and the VGG-16 verification step, triggered only on demand for candidate wrong-way trajectories, added an amortized 5 milliseconds per frame. Average alarm latency across the test set was 9.8 seconds, a modest delay the authors attribute to the secondary verification stage that buys the dramatic reduction in false alarms. In open highway scenes, the YOLOv8+ByteTrack baseline actually triggered alarms slightly faster at 7.8 seconds, but at nearly eight times the false alarm rate.

The authors are candid about limitations. The current calibration strategy handles single-roadway and two-roadway highway scenes but not complex interchanges, ramps, or urban intersections. The rotation module targets large viewpoint reversals rather than gradual drift or zoom, and the feature-difference verification may lose robustness under heavy rain, dense fog, or strong road reflections that shift the feature distribution. A single camera view also cannot fully resolve long-term occlusion or blind spots behind large vehicles. Future work, the team says, will explore weather-aware feature normalization, multi-camera cooperative detection, cross-region adaptation, and optimization for embedded edge hardware. Still, the core message stands: treating wrong-way detection as an operational reliability problem, not merely a computer vision benchmark, may be what finally makes roadside intelligent infrastructure trustworthy enough to shift highway safety from post-incident response to genuine prevention.

Subject of Research: Adaptive calibration and spatio-temporal filtering for robust camera-based highway wrong-way driving detection

Article Title: Adaptive calibration and spatio-temporal filtering for robust highway wrong-way driving detection

Article References: Liu, D., Yuan, X., Zhang, Z., Zhang, L., Wang, H., & Liu, Y. (2026). Adaptive calibration and spatio-temporal filtering for robust highway wrong-way driving detection. Frontiers in Future Transportation, 7, Article 1935771. https://doi.org/10.3389/ffutr.2026.1935771

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1935771

Keywords: wrong-way driving, highway surveillance, adaptive calibration, spatio-temporal filtering, computer vision, difference hashing, vehicle tracking, traffic safety, intelligent transportation systems, false alarm reduction, semantic segmentation, deep learning

The study, published in Frontiers in Analytical Science, addresses a persistent bottleneck in asbestos monitoring. When demolition or removal work on asbestos-containing building materials is performed improperly, fibers can leak into the surrounding air. In one survey at a demolition site, asbestos fiber concentrations at the entrance and exit of a security area reached 54 fibers per liter. Detecting such leaks early requires a rapid measurement method, but the current standard combines a short air collection time of 30 minutes or less with visual fiber counting under a phase contrast microscope, and the counting step alone takes 25 to 100 minutes depending on fiber density and analyst skill.

Visual counting is not only slow but also vulnerable to human error. According to the UK Health and Safety Executive’s analyst guide, fiber counting is subject to fatigue and proficiency-related mistakes. Analysts also tend to underestimate fiber numbers in samples with high dust density and overestimate them in cleaner samples, adding a layer of uncertainty that no amount of training fully eliminates. With the number of leak monitoring operations expected to rise alongside demolition activity, and with a limited pool of expert analysts whose training takes years, the demand for automation has become urgent.

The research team, led by Yukiko Iida of Environmental Control Center Co., Ltd. and Takashi Yamamoto of Japan’s National Institute for Environmental Studies, built their system on real-world data rather than laboratory simulations. Five laboratories belonging to the Japan Environmental Measurement and Chemical Analysis Association provided 98 phase contrast microscope images of ambient air or air collected near demolition sites. The samples were collected on mixed cellulose ester filters with 0.8 micrometer pores, made transparent with acetone vapor, and observed at 400x magnification. Expert analysts who had passed the most advanced rank of Japan’s asbestos fiber counting proficiency test manually counted the fibers in each image, following the criteria of the Asbestos Monitoring Manual: fiber length greater than 5 micrometers, width less than 3 micrometers, and an aspect ratio greater than 3.

Those expert counting results were used to annotate the images, creating 98 training datasets. The team then trained Segformer, a semantic segmentation model that uses a hierarchical Transformer encoder rather than the convolutional neural networks common in earlier segmentation models. Because Transformers can capture interactions between distant pixels, they are considered more accurate than purely convolutional approaches. In benchmark comparisons on the ADE20K dataset, Segformer achieved a mean Intersection over Union of 51.8 percent, well above the 40 to 44 percent achieved by traditional models such as FCN and DeeplabV3+. Of the 98 datasets, 77 were used to train the model and 21 were reserved for evaluation. Training took roughly 30 minutes on an NVIDIA GeForce RTX 3090 graphics processing unit, using the Dice Loss function and the Adam optimization algorithm.

The system operates in two stages. First, the AI model detects fibrous structures in the image. Second, an image processing component verifies whether each detected fiber meets the official counting criteria for length, width, and aspect ratio, judging fibers indistinguishable when overlapping shapes make accurate measurement impossible. The team evaluated performance on two levels: detection accuracy, which treats overlapping fiber clusters as single entities, and counting accuracy, which attempts to identify each individual fiber even when fibers overlap.

The results were strong on detection. The system achieved a recall of 0.90, meaning it found 90 percent of the fibers and clusters counted by expert analysts, with a precision of 0.68 and an F1 score of 0.77. Only 10 of the 97 fibers and clusters counted by experts were missed entirely, and 4 of those 10 oversights were actually detected by the AI but judged indistinguishable during the image processing stage because multiple fibers overlapped. When evaluated by counting accuracy, which is the harder task, the system achieved a recall of 0.78, a precision of 0.67, and an F1 score of 0.72, comparable to the manual counting performance of a skilled analyst.

The speed advantage was dramatic. Fiber detection took about 1.2 seconds per image on the GPU, allowing the system to scan 100 fields of view in roughly 2 minutes. The official Japanese method requires 25 to 100 minutes for the same task, making the automated system 12 to 50 times faster. This time saving does not yet include the manual work of adjusting focus and capturing images, which the team identifies as a target for future automation through synthesized depth of focus and continuous image capture as the field of view moves.

The study also revealed instructive failure modes. In some images, the system could not separate overlapping fiber clusters that an expert counted as two fibers, a limitation the team says cannot be resolved with current semantic segmentation models and requires analyst intervention. In others, the system misidentified impurities as fibers, sometimes caused by twisted filters or air bubbles introduced during slide preparation. One particularly bright fiber that an expert counted was missed by the AI, likely because the training images used positive contrast, with dark fibers against a bright background, making luminous fibers hard for the model to recognize. In one image, the system actually detected a fiber that the expert analyst overlooked because weak contrast made its length appear shorter than the 5 micrometer counting threshold, suggesting the AI could enable stricter judgment of sizes close to the criteria.

The accuracy was somewhat lower than the team’s previous work with simulated air samples, where a related model achieved a recall of 0.95 and precision of 0.91, a difference the researchers attribute to the complexity of real atmospheric samples containing particles other than asbestos fibers. Nevertheless, the team concludes that the system demonstrates reasonable accuracy on actual air samples and could prove useful for early detection and correction of asbestos fiber release during demolition work. They plan to expand the training dataset through ongoing collaboration with analytical companies and to update the fiber detection component as more advanced AI models become available, bringing fully automated asbestos monitoring closer to practical deployment at demolition and renovation sites.

Subject of Research: AI-based rapid asbestos fiber detection in phase-contrast microscope images

Article Title: Development of a rapid fiber-detection system using artificial intelligence in phase-contrast microscope images of actual atmospheric samples

Article References: Iida, Y., Yamamoto, T., Iwasaki, K., Yuki, K.-I., Kiri, K., Yamashiro, H., Toyoguchi, T., & Terazono, A. (2025). Development of a rapid fiber-detection system using artificial intelligence in phase-contrast microscope images of actual atmospheric samples. Frontiers in Analytical Science, 5, Article 1571840. https://doi.org/10.3389/frans.2025.1571840

Image Credits: AI Generated

DOI: 10.3389/frans.2025.1571840

Keywords: asbestos, artificial intelligence, SegFormer, phase-contrast microscopy, fiber detection, semantic segmentation, air quality monitoring, demolition sites, deep learning, environmental analysis, Japan, occupational health

Traditional approaches to semantic segmentation have treated the problem primarily as one of region classification, assigning a label to every pixel based on its appearance and context. Convolutional neural networks, and later transformer-based architectures, have become remarkably adept at recognizing what objects are present in a scene, but their ability to precisely delineate where one object ends and another begins has lagged behind. This is partly because standard cross-entropy loss functions, which dominate training procedures, treat all pixels equally regardless of their proximity to a boundary. A pixel deep inside a building rooftop contributes the same loss signal as a pixel sitting right at the edge between the roof and the adjacent street. As a result, networks can achieve high overall accuracy while systematically underperforming on the thin, structurally critical boundary regions that human observers use to judge segmentation quality.

The BTSR-Net framework addresses this imbalance by explicitly directing the network’s attention toward boundary neighborhoods and the spatial relationships between structures. Rather than treating boundaries as an afterthought to be cleaned up in post-processing, the architecture integrates boundary awareness into the core representation learning process. The Geometric Boundary-Distance Branch provides the network with a continuous spatial field that encodes how far each pixel lies from the nearest class transition, giving the model a geometric prior that helps it reason about where boundaries should fall even when local appearance cues are ambiguous. This kind of distance-based supervision has parallels in other areas of computer vision, where signed distance functions and boundary-aware losses have been used to sharpen edge predictions in medical imaging and natural scene understanding.

The multi-scale topology-guided structural fusion module tackles a complementary but equally important problem: maintaining the connectivity of elongated structures such as roads, rivers, and linear vegetation corridors. In remote sensing applications, a road that is correctly classified for 95 percent of its length but broken at a few critical junction points is far less useful than one that remains topologically connected throughout. Standard segmentation metrics like overall accuracy or even mean intersection-over-union do not fully capture this kind of topological failure, because a few disconnected pixels along a thin structure represent a vanishingly small fraction of total image area. The centerline Dice coefficient, or clDice, was developed specifically to measure how well tubular or elongated structures preserve their connectivity through segmentation, and its use in this study reflects a growing recognition within the remote sensing community that structural integrity matters as much as pixel-level accuracy for many downstream applications.

The dual-stream structural refinement mechanism represents a further layer of sophistication in the architecture. By confining residual corrections to a narrow band defined by the boundary-distance field, the refinement process avoids the risk of degrading already well-classified interior regions while focusing computational capacity precisely where errors are most likely and most consequential. This targeted approach echoes broader trends in deep learning toward sparse or attention-guided computation, where resources are concentrated on the most informative parts of the input rather than distributed uniformly. In the context of remote sensing, where images can be extremely large and computational efficiency is a practical concern, such selective refinement offers both accuracy and efficiency benefits.

The choice of HRNet-OCR as the backbone architecture is noteworthy in itself. HRNet maintains high-resolution representations throughout the network rather than progressively downsampling and then upsampling as many earlier architectures did, which helps preserve fine spatial detail that is essential for boundary delineation. The object-contextual representation component, originally proposed to model the semantic context of pixels by relating them to object regions rather than just local neighborhoods, adds a layer of global reasoning that helps disambiguate visually similar but semantically distinct areas. Combining these established components with the novel boundary and topology modules allows BTSR-Net to build on proven foundations while introducing targeted innovations where they are most needed.

The ISPRS Potsdam and Vaihingen datasets on which the method was evaluated are among the most widely used benchmarks in the remote sensing segmentation community, which makes the reported improvements directly comparable to a large body of prior work. Potsdam, with its dense historic urban fabric, red-roofed buildings, and complex road networks, presents particular challenges for boundary delineation because many structures share similar spectral signatures with their surroundings. Vaihingen, while also urban, features a somewhat different mix of building types, vegetation patterns, and impervious surfaces, providing a complementary test of generalization. The fact that BTSR-Net shows consistent gains across both datasets, rather than overfitting to the characteristics of one scene type, strengthens the case that the underlying design principles are broadly applicable.

The magnitude of the improvement in boundary-specific metrics deserves particular attention. On Vaihingen, the boundary intersection-over-union improved by 9.39 percentage points over the baseline, which is a substantial margin in a field where incremental gains of one or two points are often considered meaningful. This suggests that the explicit boundary-distance supervision is not merely a marginal refinement but addresses a genuine weakness in how standard architectures handle class transitions. Similarly, the connectivity improvements measured by clDice indicate that the topology-guided fusion module successfully preserves the structural relationships that matter for applications such as road network extraction, where a segmented road map must be navigable and topologically sound to be useful.

Looking at the broader context of remote sensing research, the emphasis on structure-aware segmentation reflects a maturation of the field. Early deep learning approaches to remote sensing largely imported architectures designed for natural images, achieving impressive but sometimes brittle results when applied to the unique characteristics of overhead imagery. Over time, researchers have developed increasingly specialized components that account for the particular challenges of aerial perspectives: the absence of a canonical viewing direction, the prevalence of repetitive textures, the importance of geometric relationships between objects, and the need to reason about structures that span large spatial extents. BTSR-Net’s combination of boundary awareness, topological guidance, and targeted refinement fits squarely within this trend toward domain-specific architectural innovation.

The practical implications of improved boundary and topology segmentation extend across numerous domains. In urban planning, accurate delineation of building footprints and road networks supports infrastructure assessment, property valuation, and growth modeling. In disaster response, rapidly generated and structurally reliable maps of damaged areas can guide emergency operations, where a road network that is incorrectly segmented as connected when it is actually severed could have serious consequences. In environmental monitoring, the connectivity of vegetation corridors and water bodies affects ecological assessments of habitat fragmentation and hydrological flow. In each of these cases, the quality of the segmentation output directly conditions the quality of decisions made downstream, making boundary fidelity and topological correctness not merely aesthetic concerns but functional requirements.

The computational architecture of BTSR-Net also raises interesting questions about the relationship between model complexity and interpretability. By introducing explicit geometric quantities such as boundary distance into the training objective, the method creates intermediate representations that have clear physical meaning, which can aid in understanding why the network makes particular predictions. This contrasts with purely end-to-end approaches where the internal representations remain opaque. For practitioners who need to diagnose failure cases or justify model outputs to stakeholders, having a network component that explicitly encodes distance-to-boundary provides a useful handle for interpretation and debugging.

Training such a model effectively requires careful consideration of how the different loss terms interact. The joint boundary-distance supervision must be balanced against the primary segmentation loss so that the network learns to prioritize boundary accuracy without sacrificing overall region classification. The refinement narrow-band must be wide enough to capture the full extent of boundary ambiguity but narrow enough to avoid unnecessary computation on well-classified areas. These hyperparameter choices, while not always highlighted in architectural descriptions, often determine whether a theoretically sound design translates into practical performance gains. The reported results suggest that the authors navigated these trade-offs successfully, though the sensitivity of performance to these choices remains an area where further ablation studies could provide additional insight.

The broader research community’s growing interest in boundary-aware and topology-aware segmentation is evident across multiple subfields. In medical imaging, where delineating organ boundaries and vascular networks is critical for diagnosis and treatment planning, similar principles of distance-based supervision and connectivity preservation have been developed in parallel. In autonomous driving, lane detection and road segmentation face analogous challenges in maintaining the continuity of thin structures under occlusion and lighting variation. The cross-pollination of ideas between these domains suggests that methods like BTSR-Net, which draw on concepts from multiple areas, may point toward increasingly unified approaches to structure-sensitive segmentation that transcend any single application domain.

As remote sensing data continues to grow in volume, resolution, and availability, the demand for segmentation methods that produce not just accurate but structurally meaningful outputs will only intensify. High-resolution imagery from commercial satellite constellations, drone surveys, and aerial campaigns is being generated at rates that far exceed manual annotation capacity, making automated segmentation the only scalable path to extracting actionable information. In this context, improvements in boundary quality and structural connectivity translate directly into reduced manual correction effort and more reliable downstream analysis. The work represented by BTSR-Net contributes to this broader effort by demonstrating that explicit geometric and topological reasoning, integrated into the architecture and training process, can yield measurable and meaningful improvements on standard benchmarks while remaining grounded in the practical realities of urban scene understanding.

Future directions for this line of research may include extending the boundary-distance framework to handle more complex topological structures such as intersections, loops, and branching networks, where simple connectivity metrics may not fully capture the structural relationships that matter. The integration of temporal information, using multi-date imagery to track how boundaries and structures change over time, could further enhance the utility of structure-aware segmentation for monitoring applications. Additionally, exploring how these refinement mechanisms perform across different sensor modalities, including synthetic aperture radar and hyperspectral imagery, would test the generality of the underlying principles. As the field continues to evolve, the combination of geometric priors, topological constraints, and targeted refinement exemplified by this work is likely to remain a productive direction for advancing the state of the art in remote sensing image analysis.

Subject of Research: Boundary and topology guided structural refinement for high-resolution remote sensing image segmentation

Article Title: BTSR-Net: Boundary–topology guided structural refinement for high-resolution remote sensing image segmentation

Article References: Wang, Z., Wang, W., Hu, Z., Li, Y., & Ma, K. (2026). BTSR-Net: Boundary–topology guided structural refinement for high-resolution remote sensing image segmentation. PLOS One, 21(10), e0360524. https://doi.org/10.1371/journal.pone.0360524

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360524

Keywords: remote sensing, semantic segmentation, deep learning, BTSR-Net, boundary detection, topology, HRNet-OCR, ISPRS Potsdam, ISPRS Vaihingen, urban mapping, clDice, structural connectivity

The motivation is straightforward. Histopathological examination of hematoxylin and eosin-stained slides remains the diagnostic gold standard for breast cancer, but it depends heavily on expert interpretation, and that interpretation is affected by workload, fatigue, and interobserver variability. As laboratories digitize their slides into whole-slide images, or WSIs, the opportunity for computational assistance grows, but so does the technical challenge. A single WSI is a gigapixel-scale image, far too large to feed directly into a neural network, so any practical system must break the slide into manageable pieces, analyze them, and then reassemble the results in a form a pathologist can inspect.

HistoBreastCAD addresses this in stages. First, a semantic segmentation model identifies tissue regions of interest, separating diagnostically relevant tissue from background. Second, a patch-based classifier assigns each analyzed region to one of four diagnostic categories: normal tissue, benign lesion, ductal carcinoma in situ, or invasive carcinoma. Third, the patch-level predictions are mapped back onto the original slide as color-coded probability overlays, so that model outputs can be reviewed in their spatial context. Finally, a web-based interface built on a React frontend and a FastAPI backend lets users pan and zoom across the gigapixel slide with OpenSeadragon, view the automated overlays, and draw their own manual annotations in a separate layer.

The segmentation module was trained and compared using three established architectures: U-Net with a ResNet-34 encoder, DeepLabV3 with a ResNet-50 backbone, and DeepLabV3+ with a ResNet-50 backbone. DeepLabV3+ combines atrous convolutions with atrous spatial pyramid pooling to capture contextual information at multiple receptive-field scales, while its decoder progressively recovers spatial detail. The models were trained on 256-by-256-pixel tiles extracted from ten annotated whole-slide images from the public BACH dataset, producing 248,501 tiles at native resolution and 16,421 tiles at the next downsampled pyramid level. Training used a combined Dice and focal loss with the AdamW optimizer, mixed precision, and early stopping.

Crucially, the evaluation was designed to prevent data leakage. Because multiple patches derive from the same source slide, the team grouped cross-validation folds by whole-slide image rather than by tile: in each of five folds, two complete slides were held out for validation while the remaining eight were used for training, and each slide appeared in validation exactly once. Under this protocol, the final DeepLabV3+ configuration at native resolution achieved an out-of-fold micro-Dice of 0.8390, an intersection-over-union of 0.7226, precision of 0.8278, and recall of 0.8505. DeepLabV3 performed closely behind, while U-Net showed markedly greater instability across folds, with fold-level micro-Dice standard deviations exceeding 0.28 at the downsampled level.

A subtle methodological finding emerged from the choice of checkpoint-selection metric. Tile-averaged Dice, the team found, can be misleadingly inflated by empty-background tiles that score perfectly while carrying no diagnostic information, potentially rewarding checkpoints with very low recall on actual tissue. When the researchers retrained DeepLabV3+ using pooled pixel-level micro-Dice directly for checkpoint selection, the aggregate Dice stayed essentially unchanged, but recall rose from 0.7828 to 0.8505 while precision fell from 0.9039 to 0.8278. In other words, the better metric did not manufacture a higher score; it shifted the operating point toward sensitivity and eliminated a silent failure mode.

For classification, the team evaluated four convolutional architectures on 400 labeled BACH microscopy images, 100 per diagnostic class, again with source-image-grouped five-fold cross-validation so that all 48 patches from a given source image stayed within the same partition. EfficientNetB0, which uses compound scaling of network depth, width, and input resolution, was the strongest individual model with a macro-F1 of 0.8142. The final system, however, uses a frozen equal-weight soft-voting ensemble of EfficientNetB0 and a ResNet18 variant with a global-average-pooling attention branch. The ensemble averages the two models’ softmax probability distributions with weights of 0.5 each, introducing no additional trainable parameters, and achieved an accuracy of 0.8283 and a macro-F1 of 0.8282, outperforming EfficientNetB0 alone in all five validation folds with a paired t-test p-value of 0.0134.

The most sobering results came from an exploratory reader study on the 100 microscopy images of the held-out BACH test set, which lack public reference labels. Two pathologists independently assigned four-class diagnoses, with 95 images receiving valid paired annotations. Interobserver agreement was moderate, with observed agreement of 66.32 percent and a Cohen’s kappa of 0.5105. Model-to-reader concordance was lower: EfficientNetB0 reached kappa values of 0.3242 against one reader and 0.2183 against the other, and the ensemble did not improve on this despite its benchmark advantage. Notably, the models assigned roughly 35 percent of the test images to the normal category, while each reader used that label for only 10 cases, a distributional mismatch that likely contributed to the disagreement.

The authors are careful about what these numbers mean. The cross-validated classification results quantify performance on held-out BACH source images, not end-to-end diagnostic accuracy of the complete pipeline, and the reader study measured concordance with individual expert opinions rather than correctness against an adjudicated reference standard. The system also relies on a single public benchmark, ten annotated slides for segmentation, and established CNN architectures rather than the large-scale pathology foundation models now emerging in the field. External multi-institutional validation, prospective multi-reader studies with consensus diagnoses, and formal explainability assessment are all listed as prerequisites before any claim of clinical generalizability.

Even with those caveats, the study offers a useful template for how computational pathology tools should be evaluated. By enforcing source-level partitioning, auditing tile metadata, confining augmentation and class balancing to training data, and separating benchmark performance from reader agreement, HistoBreastCAD demonstrates that an integrated segmentation-classification-visualization workflow is technically feasible with modest hardware, including consumer-grade GPUs, and that the discipline of leakage-safe evaluation can reveal exactly where a system’s strengths end and its open questions begin.

Subject of Research: Deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images

Article Title: HistoBreastCAD: an integrated deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images

Article References: Celi-Vega, O., Jiménez-Gaona, Y., Corral-Dominguez, P., Gutiérrez, Á., & Socola-Barzallo, Z. (2026). HistoBreastCAD: an integrated deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images. Frontiers in Bioinformatics, 6, Article 1870283. https://doi.org/10.3389/fbinf.2026.1870283

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1870283

Keywords: breast cancer, histopathology, deep learning, whole-slide imaging, semantic segmentation, computer-aided diagnosis, ensemble classification, EfficientNet, DeepLabV3+, BACH dataset, digital pathology, interobserver variability

In a study published in Frontiers in Manufacturing Technology, engineers from Cranfield’s Centre for Computational Engineering Sciences, the Welding and Additive Manufacturing Centre, and the Centre for Robotics and Assembly unveiled WAAM-ViD, a vision-based monitoring pipeline that can analyse the melt pool in wire-based directed energy deposition regardless of where the camera sits, what kind of camera is used, or how the lighting conditions change. The work addresses one of the most stubborn limitations in additive manufacturing quality control: the tendency of monitoring systems to be locked to a single machine, a single camera position, and a single set of optical assumptions, making them nearly impossible to transfer between industrial environments.

WAAM itself is formally classified under the ISO/ASTM 52900 standard as directed energy deposition using an arc, or DED-Arc. It works by harnessing conventional welding technology, using a plasma or electric arc torch to melt metal wire feedstock and deposit it in layers. The appeal is obvious: deposition rates are high, raw material is cheap wire rather than expensive powder, and the technique scales to components measured in metres rather than millimetres. Powder bed fusion, by contrast, offers exquisite precision for small, intricate parts but cannot match WAAM’s throughput for large structures. That is why aerospace manufacturers, who need both complexity and scale, have been among the earliest and most enthusiastic adopters of metal additive manufacturing in general and WAAM in particular.

The challenge is that WAAM is exquisitely sensitive to its own process parameters. The researchers identify three primary sources of defects: material cleanliness, where impurities in the feedstock wire promote porosity and internal flaws; material reactions, where the interaction between wire and arc produces cracking and microstructural irregularities during solidification; and process instability, which manifests as humping, spatter, overlapping, collapse, and machine suspension. Internal defects often require advanced non-destructive techniques such as computed tomography to detect after the fact. Process instabilities, however, can often be observed in real time, which is precisely where vision-based monitoring earns its keep. A camera watching the melt pool can catch anomalies the instant they form, rather than discovering them in a finished part weeks later.

Previous approaches to vision-based WAAM monitoring have achieved impressive results but with significant caveats. Some studies positioned cameras behind the welding torch to minimise arc glare, achieving defect classification accuracies above 95 percent, but this rear-mounted configuration introduces a fundamental delay: defects are only observed after deposition has already occurred. Others focused directly on the melt pool using high dynamic range cameras and narrowband optical filters to suppress the intense arc light, training deep learning models such as ResNet and ResNeXt to classify defects with accuracies approaching 99 percent. Yet when researchers attempted actual dimensional analysis of the melt pool, they ran into a wall of hardware dependency. Threshold-based segmentation, the standard technique for isolating the melt pool in an image, is notoriously sensitive to lighting variations and arc interference. Systems calibrated on one camera at one angle simply failed when moved.

The Cranfield team’s answer is a two-stage deep learning pipeline that decouples the monitoring problem from any specific imaging setup. The first stage uses DeepLabv3, a semantic segmentation architecture built on a ResNet-50 backbone, to identify and isolate the melt pool in each video frame. Rather than training on a fully labelled dataset, which would have required enormous annotation effort, the researchers employed an active learning paradigm: the model was initially trained on limited labelled data, then iteratively fine-tuned on the frames where it was least confident, as measured by Shannon entropy. Four independent raters assessed segmentation quality using a structured assessment form, and videos with the highest uncertainty and lowest qualitative scores were prioritised for annotation in subsequent training cycles. The process converged after five cycles, with the fourth cycle’s parameters delivering the best performance: a median Dice similarity coefficient of 95.90 percent, a standard metric of overlap between predicted and ground-truth segmentation masks.

The second stage is where the innovation becomes genuinely distinctive. WAAM-ViDNet, a regression-based multimodal model, takes the segmented melt pool image and fuses it with camera calibration metadata, including the camera matrix, distortion coefficients, rotation matrix, and translation vector extracted for each viewing angle. This allows the model to translate what it sees in pixel space into physically meaningful dimensions in millimetres, regardless of perspective. The architecture uses AlexNet as its feature extractor, chosen for its proven effectiveness and lightweight structure, with features concatenated with calibration data and passed through fully connected layers to predict melt pool width. Among five candidate configurations tested, the researchers selected a variant with five fully connected layers that achieved 88.71 percent accuracy while keeping computational burden low enough for real-time deployment.

To train and validate this system, the team built something the field has lacked entirely: a public benchmark dataset. WAAM-ViD comprises 110 videos with an average duration of 26.5 seconds, captured at 25 to 30 frames per second using two fundamentally different cameras. Eighty videos came from a passive Xiris XVC-1000 high dynamic range CMOS camera, which boasts a dynamic range exceeding 140 decibels and can image directly into the arc’s glare. Thirty came from a CAVITAR C400-H active laser illumination camera, which uses a narrowband pulsed laser to suppress arc light interference entirely. Cameras were positioned at five orientations relative to the plasma torch, including top-backwards, top-central, top-forward, side-central, and side-forward, with angle and distance systematically varied. Ground truth weld bead dimensions were measured post-process using a Micro-Epsilon laser profile sensor with 1.5-micrometre resolution, then temporally aligned with each video frame. The dataset, complete with manually annotated segmentation masks in COCO JSON format and full metadata, is now publicly available.

The results reveal both the promise and the honest limitations of the approach. Overall prediction accuracy of 88.71 percent is lower than the 97 to 99 percent figures reported by studies using single, optimally positioned cameras with unobstructed top-down views of the melt pool. But the Cranfield researchers argue this comparison misses the point: their system was evaluated across multiple camera viewpoints with varying occlusions, a substantially harder learning task that no prior study had attempted. Performance varied between 9.43 and 13.23 percent error depending on orientation, with the best results from the top-forward HDR configuration and the worst from side orientations with the least training data. Inference speed is remarkable: the full pipeline, segmentation plus width prediction, runs in 0.014 seconds per frame on an Nvidia RTX 4070 Ti Super GPU, fast enough for genuine real-time monitoring, though the researchers caution that performance would degrade on less powerful hardware.

The researchers are candid that their work is a foundation rather than a finished product. The current post-processing approach, which retains only the largest connected segmentation mask, risks suppressing the true melt pool if noise happens to be larger. The dataset contains only normal operating conditions, without simulated defects such as humping, and was collected on a single system with a single alloy, ER70S-6 steel. The authors also point toward Industry 5.0, arguing that explainability and trustworthiness will be essential for human-machine collaboration, and that uncertainty quantification, which they explored for active learning, deserves deeper investigation as a tool for building operator confidence. Still, the direction is clear. By demonstrating that melt pool analysis can survive changes in camera angle, quality, and illumination, and by releasing the data and code to prove it, the Cranfield team has laid the groundwork for monitoring systems that travel between factories rather than being welded, figuratively speaking, to a single machine. For a manufacturing technology whose economic case depends on producing large, safety-critical components reliably, that kind of portability may prove to be the difference between WAAM remaining a laboratory curiosity and becoming a mainstay of industrial production.

Subject of Research: Angle-invariant vision-based melt pool monitoring for wire arc additive manufacturing using deep learning

Article Title: WAAM-ViD: towards universal vision-based monitoring for wire arc additive manufacturing

Article References: Kim, K. W., Kamerkar, A., Chiu, T.-E., Abdi, I., Qin, J., Suder, W., & Asif, S. (2025). WAAM-ViD: towards universal vision-based monitoring for wire arc additive manufacturing. Frontiers in Manufacturing Technology, 5, Article 1676365. https://doi.org/10.3389/fmtec.2025.1676365

Image Credits: AI Generated

DOI: 10.3389/fmtec.2025.1676365

Keywords: wire arc additive manufacturing, WAAM, melt pool monitoring, deep learning, semantic segmentation, computer vision, directed energy deposition, additive manufacturing, benchmark dataset, active learning, Industry 4.0, quality control