The core of the team’s approach is a three-dimensional, multi-scale convolutional neural network, a type of deep learning architecture that has proven especially well suited to medical image segmentation. Unlike two-dimensional networks that analyze brain scans slice by slice, a 3D CNN processes the entire volumetric scan at once, allowing it to capture spatial context in all three anatomical planes. This matters for stroke imaging because infarcts are irregular, three-dimensional objects whose appearance can change dramatically between adjacent slices. The multi-scale design is equally important: by analyzing the image at multiple resolutions simultaneously, the network can detect both large, obvious lesions and small, subtle ones that a single-scale model might miss. The specific architecture the researchers employed, DeepMedic, uses parallel pathways, one trained on the full image and another on tightly cropped regions of interest, so that the network learns both global anatomical context and fine local detail around candidate lesions.

Before the infarct segmentation could happen, however, the system needed to solve a deceptively simple problem: isolating the brain from the rest of the head. DWI scans include the skull, eyes, and surrounding soft tissue, none of which are relevant to infarct quantification, and leaving them in can confuse downstream algorithms. The team automated this preprocessing step with a two-dimensional Mask R-CNN, a region-based convolutional neural network that first proposes candidate regions of the image and then generates pixel-level masks for each. In effect, the pipeline works in two stages: the Mask R-CNN carves out the brain parenchyma, and the 3D multi-scale CNN then hunts for ischemic lesions within that extracted tissue. The researchers also built a graphical user interface so that clinicians can run the entire workflow, from raw scan to quantified lesion volume, without writing a line of code, a practical touch that could determine whether such tools actually reach the clinic.

To develop and validate the system, the team retrospectively collected DWI data from 184 ischemic stroke patients, comprising 218 scans, at Taipei Veterans General Hospital. Retrospective designs like this one, in which the model is trained on scans that were acquired for routine clinical care, allow researchers to build large datasets without exposing patients to additional imaging. The scans were paired with expert segmentations that served as the ground truth against which the network’s output was measured. Performance was assessed using five-fold cross-validation on a validation set, a technique in which the data are split into five parts and the model is repeatedly trained and tested on different combinations, providing a robust estimate of how it will behave on unseen cases.

The results on brain extraction were striking. The Mask R-CNN stage achieved a mean Dice coefficient of 0.974, with a precision of 0.987 and a recall of 0.962. The Dice coefficient, a standard metric in segmentation research, ranges from 0 to 1 and measures the overlap between the model’s output and the reference segmentation, with 1 representing perfect agreement. A score approaching 0.97 means the automated brain extraction was nearly indistinguishable from expert work, which is critical because errors at this stage would propagate into the infarct measurements downstream.

Infarct segmentation, a harder task, still produced solid numbers. On the validation set, the 3D multi-scale CNN achieved a mean Dice coefficient of 0.746, with precision of 0.767 and recall of 0.768. For context, segmenting acute ischemic lesions is widely regarded as more challenging than segmenting organs or whole brains, because infarcts vary enormously in size, shape, and location, and DWI acquisition protocols differ between scanners and institutions. A Dice score in the mid-0.7 range places this model among the competitive results reported for this task, and the balanced precision and recall indicate the network neither systematically overestimates nor underestimates lesion extent, an important property when the measurements feed into treatment decisions.

Perhaps the most clinically significant finding came from testing the model on data it had never seen from institutions it had never encountered. The researchers evaluated the system on three external test sets: 66 scans from Landseed Hospital, 78 scans from Jen-Ai Hospital, and 59 scans from Taoyuan General Hospital. On these external datasets, the model yielded Dice coefficients of 0.729, 0.707, and 0.684, respectively. The drop from internal validation performance is expected, since different hospitals use different MRI scanners, field strengths, and imaging parameters, and this domain shift is precisely what has limited the real-world deployment of many medical AI systems. The fact that the model retained scores near 0.7 across all three external sites demonstrates robust generalizability, suggesting the pipeline could be transported to new hospitals without requiring complete retraining from scratch.

The study also uncovered an intriguing temporal effect: segmentation accuracy was significantly higher when the model was applied to MRI scans obtained four to seven days after stroke onset compared with scans taken on the day of the stroke itself. This likely reflects the biology of evolving infarcts. In the earliest hours after an arterial occlusion, the affected tissue may show only subtle and patchy diffusion changes, making the lesion harder to delineate even for expert human readers. As the infarct matures over the following days, the diffusion abnormality becomes more conspicuous and spatially coherent, giving the network a clearer target. For clinicians, this finding carries a practical implication: automated quantification is most reliable in the subacute window, while very early scans may warrant closer human review of any automated output.

The broader significance of this work lies in what automated lesion quantification could change about stroke care. Lesion volume is a key variable in treatment decisions, particularly for endovascular therapy, where the size of the established infarct core helps determine whether the brain tissue can still be saved. It is also a primary outcome measure in stroke clinical trials, where thousands of scans must be analyzed consistently. Manual delineation introduces interobserver variability that can blur both individual treatment decisions and trial results. An automated system that produces reproducible volumes in seconds could accelerate imaging workflows in stroke units, reduce the workload on neuroradiologists, and standardize measurements across centers. The authors note that their scheme could significantly accelerate stroke imaging analysis, reduce interobserver variability, and enhance clinical decision-making in the management of acute ischemic stroke.

Challenges remain before such tools become routine. The model was trained and tested on data from Taiwanese hospitals, and further validation across more diverse populations, scanner vendors, and image quality conditions would strengthen the case for widespread adoption. Very small or very early lesions, as the timing analysis suggests, remain the hardest cases. Yet the architecture and pipeline described here, combining automated brain extraction with multi-scale 3D segmentation and a clinician-friendly interface, represent a concrete step toward AI-assisted stroke imaging that works outside the single center where it was born. As deep learning systems continue to demonstrate that they can generalize across institutions, the vision of rapid, reproducible, and quantitative stroke assessment on every scan moves closer to everyday clinical reality.

Subject of Research: Automated segmentation and quantification of acute ischemic stroke lesions using 3D multi-scale convolutional neural networks

Article Title: Using 3D multi-scale convolutional neural networks to automate the segmentation and quantification of acute ischemic lesion

Article References: Chou, C.-C., Chen, M.-Y., Chiu, H.-H., Chen, P.-K., & Peng, S.-J. (2026). Using 3D multi-scale convolutional neural networks to automate the segmentation and quantification of acute ischemic lesion. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02729-x

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02729-x

Keywords: deep learning, convolutional neural network, ischemic stroke, diffusion-weighted imaging, infarct segmentation, medical imaging, artificial intelligence, brain extraction, Mask R-CNN, DeepMedic, lesion volume quantification, stroke MRI

The macular hole is a full-thickness defect in the macula, the pinprick-sized patch of retinal tissue responsible for sharp central vision. When it forms, patients typically notice distortion, a dark or missing spot in the middle of their visual field, and progressive loss of reading ability. Optical coherence tomography, or OCT, is the diagnostic workhorse: it uses low-coherence light to produce cross-sectional images of the retina with micrometer-scale resolution, letting clinicians see the hole’s edges, the lifted cuff of surrounding tissue, and the state of the underlying choroid. From these scans, ophthalmologists extract a handful of measurements — minimum hole diameter, base diameter, hole height on the left and right sides, and central choroidal thickness — that together inform staging and surgical strategy.

The problem is that these measurements are slow, subjective, and vulnerable to inter-observer variability. Two experienced graders can produce different numbers from the same scan, and in a busy clinic the differences can matter: hole size and configuration influence whether a surgeon expects a high closure rate with a simple gas tamponade or needs more extensive vitreoretinal surgery. Automating the extraction of these parameters has therefore been a long-standing goal, but medical images pose distinctive challenges for deep learning. OCT scans are grayscale, low in contrast compared with natural photographs, and dominated by layered textures that differ radically from the colorful, object-rich images that standard detection networks were designed to handle.

The research team, led by Zhiyuan Zhao, Xinqi Yu, Xiaochun Wang, Bin Wu, and Sheng Zhou from the Institute of Biomedical Engineering of the Chinese Academy of Medical Sciences and Peking Union Medical College together with Tianjin Eye Hospital, tackled these challenges with a fusion architecture. LMMR-YOLO — short for Lightweight Model of Mask R-CNN-YOLO — combines two of the most influential families in computer vision. Mask R-CNN, a two-stage detector prized for its precise instance segmentation, was used to expand the training dataset in a targeted way and to contribute fine-grained localization. YOLOv12-obb, a modern single-stage detector oriented around rotated bounding boxes, supplied the speed and efficiency backbone. The fusion means the system can both delineate the lesion accurately and detect it fast enough for routine use.

Several technical innovations underpin the model’s performance on grayscale medical data. The researchers integrated a grayscale image adaptation module with Efficient Channel Attention, a mechanism that lets the network learn which feature channels carry the most diagnostic information — essentially teaching it where to look in the layered structure of a retinal scan. They replaced heavy backbone components with C2f_Ghost and A2C2f_Lite modules, lightweight building blocks that preserve feature extraction power while slashing computational cost. An Oriented Bounding Box detection head allows the model to fit tightly around lesions that appear at arbitrary angles in the scan, and an Image Enhancement Algorithm sharpens contrast before detection, with a Fusion Detection Algorithm merging the outputs of the enhanced pipeline and the Mask R-CNN branch into a single verdict.

The study was retrospective, drawing on 606 OCT images from 61 patients, with ethics approval from Tianjin Eye Hospital and a waiver of individual informed consent because the data were anonymized. Performance was evaluated with the standard battery of detection metrics: precision, recall, mean average precision at an intersection-over-union threshold of 0.5, F1-score, and accuracy. On the internal test set, LMMR-YOLO achieved an mAP@0.5 of 92.69 percent and an overall accuracy of 94.78 percent for key-point detection of macular hole parameters. Precision reached 97.14 percent, meaning that when the model flags a measurement point, it is almost always correct, while recall of 87.88 percent indicates it catches the large majority of true points. The harmonic F1-score of 92.28 percent balances those two figures and sits comfortably in the range clinicians would expect from careful human graders.

Two secondary results are particularly telling about the model’s reliability. Normalized classification accuracy for central choroidal thickness — a parameter that reflects the health of the vascular layer beneath the retina and is increasingly used in staging — came out at 0.97. And the mirror misidentification rate between left hole height and right hole height, an error mode in which a system confuses the two sides of an asymmetric hole, was just 0.03. Those numbers matter because staging systems for macular holes depend on exactly such lateral distinctions; a model that routinely swapped left and right would be clinically useless no matter how impressive its headline accuracy.

Equally important is what the model does not cost. LMMR-YOLO carries only 3.34 million trainable parameters and requires 11.10 gigafloating-point operations per inference — figures that place it firmly in the lightweight category, small enough to contemplate running on standard clinic hardware or even portable devices rather than server farms. Inference speed measured 9.44 frames per second, approaching real-time performance, and the training time was short. In medical AI, where many celebrated models demand billions of parameters and hours of GPU time per scan batch, a system that balances accuracy against efficiency is arguably more valuable than a marginally more accurate but computationally prohibitive alternative. The authors frame this balance as a new paradigm for medical grayscale image analysis and model training.

The clinical implications extend beyond the numbers. Accurate, automated extraction of minimum hole diameter, base diameter, hole height, and choroidal thickness could standardize staging of full-thickness and lamellar macular holes, reduce grading variability between centers, and support personalized surgical planning — for example, helping surgeons anticipate closure likelihood and choose among operative techniques. Because the pipeline is fast and light, it could also be embedded in screening workflows, flagging urgent cases in high-volume imaging services where specialist reading time is scarce. The researchers suggest the approach supports ophthalmic diagnosis, staging, and personalized care, and its architecture — grayscale adaptation, channel attention, oriented detection, and fusion of segmentation and detection branches — is general enough to be adapted to other retinal and medical imaging tasks.

Caveats remain, as they do for any retrospective single-center study. The dataset of 606 images from 61 patients, while substantial for a rare lesion, will need validation on larger and more diverse cohorts before the model can be trusted across different scanner manufacturers, image qualities, and patient populations. The published version is an early-release, peer-reviewed accepted manuscript that will undergo further editorial refinement. Still, the trajectory is clear: the fusion of segmentation precision with detection speed, tuned specifically for the grayscale, low-contrast world of OCT, has produced a system that measures the anatomy of a sight-threatening lesion nearly as well as experts do — and does it in milliseconds. For the millions of people at risk of macular holes worldwide, that speed and consistency could eventually translate into earlier detection, better-planned surgery, and more preserved vision.

Subject of Research: Deep learning-based automatic detection and measurement of macular hole parameters in optical coherence tomography images

Article Title: A deep learning-based study on automatic extraction and measurement of macular hole parameters

Article References: Zhao, Z., Yu, X., Wang, X., Lu, K., Wu, B., & Zhou, S. (2026). A deep learning-based study on automatic extraction and measurement of macular hole parameters. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02899-8

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02899-8

Keywords: macular hole, deep learning, optical coherence tomography, LMMR-YOLO, Mask R-CNN, YOLOv12, medical imaging, ophthalmology, retinal disease, image segmentation, lightweight neural network, computer-aided diagnosis

Temperate savanna woodlands form a broad ecotone between closed forests and open grasslands, and the ones studied here span two vast sandy regions of Inner Mongolia: the Hunsandak Sandland, covering roughly 52,000 square kilometers, and the Horqin Sandland, extending over about 35,100 square kilometers. These landscapes support exceptionally rich biodiversity and act as reservoirs of genetic resources, yet the woody vegetation that anchors them has received far less attention than the grasses and herbs around it. Because trees and shrubs in these systems control carbon storage, community stability, and the spatial distribution of biodiversity, knowing exactly where each individual tree stands, and how large its crown is, is fundamental to nearly every ecological question that can be asked about the region.

The challenge is that savanna trees are maddeningly heterogeneous. Natural savanna woodlands stretch across large environmental gradients, and the woody vegetation responds with pronounced spatial variation, producing wildly diverse phenotypes at the level of single trees. Classical image-processing techniques such as local maximum filtering, marker-controlled watershed segmentation, template matching, region growing, and edge detection can work under simple conditions, but they falter in natural forests with complex canopy structures. They depend on manual parameter tuning, extract features only crudely, and generalize poorly. Deep convolutional neural networks offered a way forward, but the standard workhorse for instance segmentation, Mask R-CNN, carries a heavy price: its ResNet-50 backbone and Feature Pyramid Network demand substantial computational resources, and the pyramid module can lose information when propagating features to high resolutions, undermining the detection of small or sparsely distributed crowns.

To build DeepTree, the research team first assembled an unusually rich dataset. Between July 20 and 30, 2020, a DJI Phantom 4 RTK quadcopter flew nine sample plots across six banners of Inner Mongolia under clear, cloudless skies, capturing thousands of photographs at roughly 100 meters altitude with 75 percent forward and 70 percent side overlap. The imagery was processed into nine digital orthophotos with pixel resolutions between 2.63 and 3.06 centimeters, then resampled to a uniform 10 centimeters and sliced into 1,000-by-1,000-pixel tiles, each covering one hectare, yielding 663 tiles in total. From eight of the plots, the team manually annotated 8,724 individual tree crowns using Labelme, exporting the labels in COCO format. The ninth plot was deliberately held back as an independent cross-site test, ensuring that no spatially overlapping regions leaked between training and evaluation data.

The architectural surgery at the heart of DeepTree is elegant in its simplicity. The researchers replaced ResNet-50 with MobileNet V3, a backbone that uses just 13 layers and depthwise separable convolutions instead of 50 conventional layers, dramatically cutting computational complexity. They swapped the Feature Pyramid Network for NAS-FPN, a feature-fusion module discovered through neural architecture search, which enhances multi-scale representation and improves detection of small targets while using only 32 channels. The framework then branches into two variants: MNMS R-CNN (S), which pairs this lightweight backbone and neck with a standard region-of-interest head for maximum speed, and MNMS R-CNN (L), which adds an enhanced SCNet-based head with Squeeze-and-Excitation, Global Context, and Feature Relay components to prioritize detection completeness and segmentation accuracy.

The performance gains are remarkable. MNMS R-CNN (S) slashed computational cost from 248.0 GFLOPs to just 5.6 GFLOPs and shrank the parameter count from 42.9 million to 3.8 million, while boosting inference speed from 10.2 to 19.2 frames per second, an 88.2 percent improvement, all on 1.5 gigabytes of GPU memory instead of 6.2. The larger variant, MNMS R-CNN (L), required only 9.0 GFLOPs and 8.1 million parameters yet achieved the best overall accuracy in the study: bounding-box average precision rose from 34.9 to 46.5 percent, a relative improvement of about 33 percent, while recall jumped from 62.4 to 81.2 percent and the F1 score climbed to 81.1 percent. Comparisons against YOLO11n-seg and the high-capacity HTC framework confirmed that neither compact size nor raw capacity alone delivers the right balance; MNMS R-CNN (L) outperformed both on the key detection metrics while remaining far leaner than HTC’s 76.93 million parameters and 401.1 GFLOPs.

Perhaps the most ecologically telling results emerged when the models were tested across different canopy densities. In high-density scenes, MNMS R-CNN (L) lifted recall from Mask R-CNN’s 45.5 percent to 84.9 percent, cutting missed crowns from 72 to just 20 in a representative case. Even in sparse, low-density stands, where isolated small crowns blend into the grassy background, it nearly doubled recall, from 28.2 to 46.5 percent. Stratified analysis revealed a persistent pattern: omission errors concentrated among smaller crowns, particularly in low-density scenes where weak spectral and textural contrast against grass undermined detection. The team also found that performance degraded as imagery was downsampled, with F1 falling from 81.1 percent at 10 centimeters to 74.1 percent at 0.5 meters and 55.4 percent at 1.0 meter, underscoring how much fine spatial detail matters for resolving individual crowns.

Beyond simply finding trees, DeepTree extracts the structural traits that ecologists actually need. For crown width, the coefficient of determination reached 0.98 with a root mean square error of just 0.30 meters; for crown area, R squared was again 0.98 with an error of 2.66 square meters. The model recovered crown-width estimates spanning 1.41 to 16.05 meters and crown areas from 1.33 to 203.71 square meters, closely matching reference ranges that Mask R-CNN systematically truncated. An independent field validation using 166 measured elm trees, matched to drone detections by geographic coordinates, yielded R squared values of 0.82 for crown width and 0.78 for tree height, with the modest gap attributable to the one-year interval between imagery acquisition and field measurement, GPS positioning uncertainty, and differences in how crown width is defined on the ground versus in a segmentation mask.

Applied across all nine plots, the model mapped tree densities ranging from about 30 to 102.7 trees per hectare and canopy areas from roughly 709 to 3,117 square meters per hectare, revealing stark spatial heterogeneity in woody cover. In the held-out ninth plot, never seen during training, the model still achieved 75.8 percent precision and 76.0 percent recall, demonstrating genuine cross-site transferability, the property that separates a local demonstration from a scalable monitoring tool. The authors acknowledge limitations: RGB imagery alone struggles when crowns, shrubs, and grass share similar colors, and future work should fuse in LiDAR and spectral data, expand training libraries to more species and seasons, and test the framework across broader environmental gradients. Even so, DeepTree points toward a future where drone-based biodiversity assessment, carbon accounting, and long-term monitoring of the world’s fragile forest-grassland ecotones can run on modest hardware, one tree at a time.

Subject of Research: Lightweight deep learning instance segmentation for individual tree crown delineation and structural trait extraction from UAV imagery in temperate savanna woodlands

Article Title: DeepTree: Deep learning-based lightweight instance segmentation for individual tree crown delineation and structural trait extraction

Article References: Duan, T., Yang, B., Li, X., Hou, D., Hu, P., Li, X., Cong, W., & Wang, F. (2026). DeepTree: Deep learning-based lightweight instance segmentation for individual tree crown delineation and structural trait extraction. Smart Agricultural Technology, 15, Article 102552. https://doi.org/10.1016/j.atech.2026.102552

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102552

Keywords: deep learning, UAV remote sensing, instance segmentation, tree crown delineation, Mask R-CNN, MobileNet V3, NAS-FPN, savanna woodlands, Inner Mongolia, ecological monitoring, phenotype extraction, lightweight models