The retrospective study, approved by the hospital’s institutional review board with informed consent waived and all patient data anonymized, set out to build models that could classify disease activity in patients with small bowel Crohn’s disease, a form of the condition that is notoriously difficult to monitor because the affected segments often lie beyond the reach of conventional colonoscopy. The team, led by Yi Shen, Chang Rong and Han Zhang from the Department of Radiology together with gastroenterologist Jing Hu and corresponding author Xingwang Wu, drew on three complementary data streams collected as part of routine clinical workup, then asked a simple question with far-reaching implications: does a model that sees everything outperform models that see only one thing?

The first stream was clinical. Using univariate and multivariate logistic regression, the researchers sifted through demographic and laboratory variables to identify which factors independently predicted disease activity. The abbreviations catalogued in the paper read like a standard inflammatory workup: C-reactive protein, albumin, fibrinogen, hemoglobin, platelet count, neutrophil percentage, fecal calprotectin and fecal occult blood, alongside the Harvey–Bradshaw Index, a bedside clinical score. Four clinical factors survived the selection process and earned a place in the final models, providing the systemic context that imaging alone cannot supply.

The second stream came from computed tomography enterography, a specialized CT technique in which patients drink neutral oral contrast to distend the small bowel, allowing radiologists to inspect the bowel wall for signs of inflammation. Rather than relying on subjective visual reads, the team applied radiomics, the practice of extracting large numbers of quantitative features from medical images, such as texture, intensity and shape descriptors, that are invisible to the human eye. Lesions were outlined using semi-automatic segmentation, and the resulting feature set was pruned with LASSO regression, a statistical technique that shrinks most coefficients to zero and keeps only the most informative predictors. Seven radiomic features made the final cut, each one a compressed mathematical summary of how inflamed tissue looks to a scanner.

The third and most novel stream was the endoscopic image itself. Double-balloon enteroscopy, a technique that uses a balloon-equipped overtube to advance a scope deep into the small bowel, produces images of mucosal lesions that no other modality can capture. The researchers trained ResNet-18, a widely used convolutional neural network architecture built around residual connections that allow very deep networks to train stably, to classify these images and to extract features from their internal layers. From this pipeline they obtained two kinds of output: a binary deep learning classification result, essentially the network’s own verdict on disease activity, and six deep learning-derived features, numerical fingerprints of the endoscopic images that carry information beyond a simple yes-or-no label.

With ingredients in hand, the team assembled seven models spanning the spectrum from single modality to full integration, and the results were unambiguous. Every model achieved an area under the receiver operating characteristic curve, or AUC, exceeding 0.90 in the training cohort, a threshold generally regarded as excellent discrimination. But the multimodal models, designated Combined 2 and Combined 3, which fused clinical data, radiomics and deep learning outputs, delivered the highest performance of all, clearly outperforming both single-modality models and models limited to clinical plus radiomic data. In other words, each additional data stream added genuine predictive value rather than redundant noise.

Crucially, the authors did not rest on AUC alone. They deployed a battery of statistical tests to confirm that the multimodal advantage was real. DeLong tests, the standard method for comparing correlated ROC curves, showed the differences between models were statistically significant. Net reclassification improvement and integrated discrimination improvement, two metrics designed to quantify how much better a new model assigns patients to the correct risk category, both confirmed the superiority of the integrated approach. Calibration analysis checked that predicted probabilities matched observed outcomes, and decision curve analysis evaluated the clinical usefulness of the models across a range of risk thresholds, asking whether acting on the model’s predictions would yield more benefit than treating everyone or no one.

Transparency was another priority. The researchers packaged their best models into nomograms, graphical calculators that let a clinician plot an individual patient’s values on each predictor axis and read off a personalized probability of active disease. To open the black box, they used SHAP, or Shapley Additive Explanations, a technique borrowed from cooperative game theory that assigns each feature a contribution to every individual prediction. The SHAP plots highlighted which clinical, radiomic and deep learning features drove the model’s decisions, offering a measure of interpretability that is increasingly demanded of medical artificial intelligence systems.

The implications extend beyond one hospital in Hefei. Crohn’s disease is a chronic, relapsing condition in which treatment decisions hinge on accurately distinguishing active inflammation from fibrotic scarring or remission, and endoscopic healing has become the gold standard therapeutic target. Yet routine endoscopic assessment of the small bowel requires double-balloon enteroscopy, an invasive, time-consuming and technically demanding procedure, while magnetic resonance and CT enterography are interpreted subjectively and vary between readers. A validated multimodal model could, in principle, let clinicians combine information they already collect, blood tests, a CT scan and endoscopic images, into a quantitative, reproducible activity score, potentially reducing unnecessary procedures and supporting earlier treatment adjustments. The authors are careful to frame the work as complementary to clinical evaluation rather than a replacement for it, and the study’s retrospective, single-center design means prospective validation in larger and more diverse cohorts will be needed before such tools enter practice.

Still, the study offers a compelling demonstration of where medical imaging and artificial intelligence are heading. Rather than betting on a single algorithm or a single modality, it treats the patient as a multimodal data object, with each examination contributing a different slice of the biological truth about inflammation. The fact that models built from routinely acquired clinical, radiological and endoscopic data could all clear the 0.90 AUC bar, and that integration pushed performance higher still, suggests that the raw material for precision medicine in inflammatory bowel disease may already be sitting in hospital archives. As deep learning feature extraction matures and explainability tools like SHAP become standard practice, the gap between research models like this one and bedside decision support continues to narrow, and for patients with small bowel Crohn’s disease, that narrowing could mean fewer invasive exams and better-timed therapy.

Subject of Research: Multimodal machine learning assessment of small bowel Crohn's disease activity using clinical data, CTE radiomics and deep learning on double-balloon enteroscopy images

Article Title: Multimodal evaluation of small bowel crohn’s disease activity integrating clinical data, CTE radiomics, and deep learning on DBE

Article References: Multimodal evaluation of small bowel crohn’s disease activity integrating clinical data, CTE radiomics, and deep learning on DBE. (n.d.). https://doi.org/10.1186/s12880-026-02878-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02878-z

Keywords: Crohn's disease, small bowel, computed tomography enterography, radiomics, deep learning, double-balloon enteroscopy, ResNet-18, SHAP, nomogram, inflammatory bowel disease, medical imaging, machine learning

The core insight behind the work is that classification errors across classes can conflict. A network that excels at distinguishing shirts from coats in the Fashion MNIST dataset might struggle with sneakers versus sandals, and the conventional practice of averaging errors into a single scalar loss hides these tensions entirely. By contrast, the CMaO-SL formulation defines one objective per class, so a ten-class problem becomes a ten-objective optimization problem. Because no single architecture can simultaneously minimize every class-specific error, the search targets a Pareto set: a collection of non-dominated solutions in which improving performance on one class necessarily comes at some cost to another. This trade-off information, normally invisible to practitioners, becomes an explicit output of the training process.

Solving such a problem is computationally daunting. Not only must the search evaluate many candidate architectures, but it must do so repeatedly to draw statistically meaningful conclusions about a stochastic process. The authors tackled this challenge with a two-pronged strategy combining transfer learning with evolutionary neural architecture search. Every network in their framework is split into two parts: a deep, pre-trained feature extractor kept completely frozen, and a shallower classification sub-network that is allowed to evolve. The feature extractor is derived from a ResNet-18 network pre-trained on ImageNet, with its final layers removed and its pooling adjusted to output 2,304 features. Because the backbone is never retrained or fine-tuned, the expensive part of the computation is paid only once, and the search can concentrate on the smaller, cheaper classification head.

The evolutionary search itself adapts MOEA/D, a well-established decomposition-based multiobjective evolutionary algorithm, to the architecture search setting. The user first defines a master architecture for the classification sub-network, consisting of fully connected layers with batch normalization and ReLU activations. Each individual in the evolving population is a binary mask vector that dictates which connections in the master architecture are pruned; where a mask entry is zero, the corresponding connection is removed and its weight fixed at zero. This pruning-as-search strategy generates a rich family of alternative architectures from a single blueprint. Each candidate is then trained with gradient descent, evaluated per class on a validation set, and assigned a performance vector whose components are one minus the F1 score for each class.

The evaluation procedure embeds an early-stopping mechanism to guard against overfitting. During each candidate’s five training epochs, the algorithm tracks a scalar validation error and retains the best-performing weight vector, ensuring that later epochs cannot degrade the network that gets assessed. The authors used two training loss functions in their experiments, cross-entropy and a weighted average of per-class binary cross-entropy aligned with each sub-problem’s reference direction, and two corresponding validation error functions. This produced five distinct search configurations, three evolutionary and two non-evolutionary, the latter serving as controls in which independent scalar searches replaced the coupled evolutionary process.

The experimental study spanned six benchmark datasets drawn from very different domains: Fashion MNIST, K-MNIST, CIFAR10, and three medical image collections from the MedMNIST suite, namely BloodMNIST, DermaMNIST and OCTMNIST. Each configuration was run fifteen times, allowing a rigorous statistical treatment that is rare in the Pareto-based architecture search literature. After confirming with a Shapiro-Wilks test that the results were not normally distributed, the authors compared configurations using medians and pairwise Mann-Whitney-Wilcoxon tests at a significance level of 0.05, alongside the IGD+ performance indicator, which measures how well an obtained front approximates a reference Pareto front.

The results were consistent across datasets and evaluation sets. Networks trained with cross-entropy loss produced better-approximated Pareto fronts than those trained with the weighted binary cross-entropy variant, and the evolutionary searches matched or beat their non-evolutionary counterparts on front quality while producing sets of networks with lower total complexity. Parallel coordinate plots of per-class F1 scores made the conflicting nature of the objectives vividly visible: the line connecting a classifier’s best performance on one class typically dipped on others. Crucially, when the authors picked the single best network from each evolutionary run’s set, measured by overall validation accuracy, those networks matched the accuracy of the unpruned master architecture while being substantially smaller, demonstrating that the many-objective formulation finds leaner models without sacrificing performance.

The benchmark comparisons are perhaps the most striking part of the study. When the networks discovered by the proposed method were plotted against published reference results for ResNet-18-based classifiers on the same datasets, no published result dominated the new networks on the combined axes of accuracy and complexity. The full ResNet-18 carries roughly 11.5 million parameters relevant to the comparison, whereas the evolved networks came in at around 2.5 million, meaning the reference architecture is about 360 percent more complex. That dramatic reduction in size cost less than 8 percent in accuracy, and on the DermaMNIST skin-lesion dataset the evolved classifiers actually surpassed the published benchmark. For a method that never fine-tuned its feature extractor and trained each candidate for only five epochs, this is a remarkable showing.

The practical implications extend beyond raw benchmark numbers. Because the output of the search is a set of classifiers with exposed trade-offs, users gain flexibility that single-model pipelines cannot offer. A deployment with tight computational budgets can select the smallest acceptable network from the set, while an application where a particular class matters most, say a rare disease category in a medical screening tool, can pick the member that favors that class. Alternatively, the trade-off structure can guide the construction of ensembles, an avenue the authors identify as a priority for future work. They also flag several open questions: whether fine-tuning the shared feature extractor would push accuracy higher, how the approach handles imbalanced data and noisy labels common in real-world problems, and how alternative pruning schemes, such as removing whole layers or blocks, might perform.

The study also fills a conceptual gap in the literature. Although Pareto-based multiobjective machine learning has existed for years, formal treatments of many-objective supervised learning, where the number of objectives exceeds three, have been scarce, and no prior work had combined such a formulation with neural architecture search and a transferred feature extractor. Most existing multiobjective NAS studies juggle only two or three objectives, typically accuracy, model size and inference cost, rather than one objective per class. By demonstrating a working pipeline at ten objectives, backed by fifteen-run statistical analysis per configuration, the Tel-Aviv and Braude team has shown that class-aware, many-objective architecture search is not only feasible but competitive. As automated machine learning matures, approaches like this suggest a future in which the question is no longer simply how accurate a network is, but which compromises it makes, and whether those are the compromises its users actually want.

Subject of Research: Class-based many-objective neural architecture search for deep classifiers using evolutionary optimization and transfer learning

Article Title: Class-based many-objective neural architecture search for deep classifiers

Article References: Mardiks, G., Salih, A., & Moshaiov, A. (2026). Class-based many-objective neural architecture search for deep classifiers. Neural Computing and Applications, 38(19), Article 800. https://doi.org/10.1007/s00521-026-12510-4

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12510-4

Keywords: neural architecture search, many-objective optimization, Pareto front, evolutionary algorithms, transfer learning, deep learning, multi-class classification, MOEA/D, pruning, ResNet-18, AutoML, machine learning

The core problem the researchers, led by Shalini Saravanan and colleagues, set out to address is the notorious unreliability of softmax confidence. Most classifiers convert their internal scores into probabilities through a softmax function, but these numbers reflect the relative magnitude of scores rather than genuine predictive certainty. The result is a well-documented pathology: softmax outputs are poorly calibrated and routinely assign high confidence to incorrect predictions or to inputs drawn from entirely outside the training distribution. Existing fixes, including Bayesian neural networks, deep ensembles, and temperature scaling, either demand substantial extra computation or require specialised training procedures that cannot be bolted onto a model already in service.

The team’s answer, which they call the class-conditioned neutrosophic representation, or CCNR, works entirely after the fact. Deep networks transform images into embeddings, high-dimensional vectors in which samples of the same class cluster together while different classes occupy distinct regions. CCNR exploits this geometry: for each embedding, it computes distances to the nearest and second-nearest class centroids and converts those distances into a triplet of values drawn from neutrosophic set theory. Truth measures the evidence supporting the predicted class, falsity measures the strongest competing evidence, and indeterminacy captures the ambiguity between them. Unlike probabilities, these three components need not sum to one, allowing the representation to express conflicting evidence explicitly rather than collapsing it into a single scalar.

The construction has elegant mathematical properties that the authors prove formally. Indeterminacy peaks exactly at the decision boundary, where a sample lies equidistant between two class centroids, and falls monotonically as the classification margin grows. The mapping is continuous and bounded, and it is provably not equivalent to softmax confidence. Two distance formulations are explored: a Euclidean version that assumes isotropic class structure, and a Mahalanobis version that incorporates the covariance of the feature space. The resulting triplet is simply appended to the original 512-dimensional ResNet-18 embedding, producing a 515-dimensional vector that feeds standard classifiers such as Random Forests, Logistic Regression, or Support Vector Machines, with six named variants covering the combinations.

Evaluation was deliberately demanding. The framework was tested under a leave-one-class-out protocol across four datasets spanning CIFAR-10 and three real-world plant disease collections, with one class withheld from training and used exclusively to test whether the model could detect genuinely novel inputs. All results were averaged over ten independent trials with fixed seeds and assessed with Holm-corrected Wilcoxon signed-rank tests. The most dramatic finding came from the Eggplant Leaf Disease dataset, where the fine-tuned ResNet-18 baseline, despite achieving 99 percent classification accuracy, collapsed to an out-of-distribution detection score of 0.366, well below random chance. The same softmax model that was nearly perfect on familiar data assigned high confidence to images of a disease it had never encountered.

Perhaps the study’s most consequential insight concerns not what the augmentation adds but what conventional readouts discard. Under the Mahalanobis formulation, the truth component is a monotone transform of the distance to the nearest centroid, meaning classical Mahalanobis out-of-distribution scoring is recovered exactly by reading that single coordinate directly. When the researchers replaced the standard uncertainty signal, one minus the maximum class probability, with the truth component, the below-chance detection scores of the linear variants on Eggplant jumped from roughly 0.46 to 0.96 in every single trial. The geometric information had been present all along; the classifier’s posterior simply threw it away. This failure, the authors stress, is a property of the readout rather than the representation, and it can be corrected without any retraining.

The team also displayed unusual scientific candour by subjecting their own method to a six-arm ladder of random-dimension controls, in which the three added columns were replaced with everything from plain Gaussian noise to the real components permuted within class. Pooled across datasets and classifiers, the augmentation was not distinguishable from matched random dimensions; only one configuration, CCNR-SVM-E on the most separable dataset, exceeded every control after correction, and by a margin of just 0.002. The authors report this boundary in place of the aggregate gains it does not support, a level of methodological honesty that is rare in the uncertainty estimation literature and that they argue should become standard practice.

Against the main computational alternative, a snapshot ensemble that saves multiple checkpoints during training, CCNR held its own on uncertainty quality, with nine of twelve comparisons statistically indistinguishable, at roughly a third of the training time, storage, and inference latency. Where a fine-tuned backbone already exists, the augmentation adds only seconds of classifier training and a few megabytes of storage. Yet the study is equally clear about limitations. Detection of corruption-based covariate shift, such as blur, brightness changes, or JPEG compression, was markedly weaker than detection of novel classes, and each feature space proved effectively blind to at least one common corruption while accuracy silently fell by 9 to 13 points. Strikingly, the choice of which class was withheld moved detection scores by 0.135 to 0.269, far more than any difference between methods.

For practitioners, the message is nuanced but actionable. The neutrosophic components carry genuine class information, classifying above chance on every dataset when used alone, with the strongest signal precisely where class clusters are geometrically clearest. A high indeterminacy score offers a natural trigger for human review, aligning with how expert agronomists already treat ambiguous symptoms. But the authors recommend pairing the method with explicit image-quality checks in uncontrolled field conditions, selecting the uncertainty readout to match the operational objective, and reporting protocol-sensitivity analyses routinely. In a field often criticised for overclaiming, this work stands out both for what it demonstrates, that feature-space geometry holds uncertainty information our standard readouts waste, and for showing exactly where its own promise ends.

Subject of Research: Uncertainty-aware classification of deep feature embeddings using class-conditioned neutrosophic representations

Article Title: Uncertainty-aware deep feature classification using class-conditioned neutrosophic representation

Article References: Saravanan, S., Kumar, K., Obbineni, J. M., & Kandasamy, I. (2026). Uncertainty-aware deep feature classification using class-conditioned neutrosophic representation. Machine Learning with Applications, 26, Article 101037. https://doi.org/10.1016/j.mlwa.2026.101037

Image Credits: AI Generated

DOI: Not provided

Keywords: deep learning, uncertainty estimation, neutrosophic sets, out-of-distribution detection, plant disease detection, ResNet-18, feature embeddings, softmax calibration, Mahalanobis distance, machine learning, precision agriculture, interpretable AI

The stakes are considerable. When a diagnostic network has memorized a patient’s histology slide or a fraud detector has internalized a transaction history, a successful membership inference can reveal that the record existed in the training set at all, exposing sensitive attributes and confidential data. Prior attacks have generally fallen into two camps. Static methods, such as thresholding the prediction loss as proposed by Yeom and colleagues in 2018, or label-aware entropy introduced by Song and Mittal in 2021, are computationally cheap but extract signals from a single model output, ignoring how predictions evolve when the input is nudged. Shadow-model methods, including the widely used LiRA attack from Carlini and colleagues, train fleets of reference models to mimic the target’s behavior, achieving strong accuracy but at a computational cost that scales linearly with the number of shadow models and degrades when data distributions shift.

PUC takes a third path. Instead of asking what the model says about a sample, it asks how the model’s answer changes as the sample is systematically perturbed. The framework integrates three complementary perturbation strategies: Temperature Scaling, which rescales the raw logits fed into the softmax function by a scalar temperature parameter and traces how confidence softens across a predefined temperature grid; Monte Carlo Dropout, which keeps dropout layers active during inference and averages confidence over multiple stochastic forward passes at varying dropout rates; and the Fast Gradient Sign Method, which pushes the input along the sign of the loss gradient and tracks how confidence decays as adversarial strength increases. Each strategy produces a continuous curve of confidence against perturbation intensity, and together they probe the model from output calibration, internal stochasticity, and local decision-boundary sensitivity angles simultaneously.

The physical intuition behind the attack is that memorization leaves a structural fingerprint. Samples that were part of the training set tend to sit in flatter, more stable regions of the model’s response landscape, so their confidence curves remain relatively smooth and flat under perturbation. Non-member samples, by contrast, typically show steeper declines in confidence as perturbations intensify. Earlier work hinted at pieces of this picture: Ravikumar and colleagues showed in 2024 that training samples occupy regions with lower Hessian traces, Rahimian and colleagues observed in 2021 that member labels are more stable under slight perturbations, and Grosso and colleagues measured the perturbation magnitude needed to flip predictions. What those approaches lacked was a unified, function-level treatment that models the entire relationship between perturbation strength and model response rather than a single geometric property.

To turn curves into decisions, the researchers built a feature engineering pipeline that compresses each continuous response curve into a compact 19-dimensional vector. Six geometric features capture the curve’s shape, including arc length, maximum and mean curvature, and the number of turning points, while thirteen time-domain statistical features quantify the distribution of confidence values along the curve, including the integral area, total variation, mean, variance, skewness, and kurtosis. Because the three perturbation families each yield their own feature set, an attention mechanism dynamically weights the three channels according to each sample’s response pattern before concatenation, producing a 57-dimensional fused representation. This avoids the information loss of equal weighting and feeds supervised classifiers: logistic regression in the PUC-LR variant, a multilayer perceptron in PUC-MLP, and LightGBM in PUC-Tree.

The experimental results are striking. On CIFAR-10 with a ResNet-18 target, PUC-Tree achieved an area under the ROC curve of 0.869, exceeding LiRA by more than 16 percentage points, while PUC-MLP reached 0.837. The advantage widens dramatically in the regime that matters most for realistic auditing and attacks: extremely low false positive rates. At a false positive rate of 0.1 percent, PUC-Tree still identified 21.1 percent of members, compared with 7.6 percent for LiRA and 3.6 percent for the curvature-based Curv method, while the loss-thresholding approach of Yeom and colleagues failed entirely. At a false positive rate of 0.01 percent, PUC-Tree identified over 20 percent of members while competing methods degraded to near-random guessing. To reach a true positive rate of 50 percent, PUC-Tree required a false positive rate of only 3.24 percent, whereas LiRA needed 26.7 percent, an eightfold difference in cost.

Generalization results reinforce the picture. On the harder CIFAR-100 benchmark, PUC-Tree and PUC-MLP reached AUCs of 0.978 and 0.974 respectively, and even the linear PUC-LR variant improved from 0.725 on CIFAR-10 to 0.918. The framework also proved architecture-agnostic, performing consistently across ResNet-18, the lightweight MobileNetV2, and the Vision Transformer ViT-B/16. MobileNetV2 posted an AUC of 0.945 on PathMNIST, a medical histology dataset of 107,180 patches spanning nine tissue types, confirming that membership inference poses a severe threat even for resource-constrained edge deployments. ViT-B/16 achieved an AUC of 0.929 on CIFAR-100 and 0.901 on PathMNIST, although its low-FPR performance on CIFAR-10 was dampened by the smoother local response to perturbations that Transformers exhibit.

Ablation studies clarified which ingredients carry the weight. Feature importance rankings from the LightGBM classifier placed the geometric attributes arc length and maximum curvature at the top across all datasets, confirming that curve trajectories carry stronger membership signals than static confidence values, with total variation and skewness leading among the statistical features. Pruning to the top 30 of 57 features retained more than 97 percent of full performance, and the top 20 features still yielded an AUC near 0.80. When perturbation combinations were varied, the full tri-perturbation setup consistently beat every single or pairwise subset, and the improvement was synergistic rather than a linear superposition, indicating that the three mechanisms capture complementary structure in memorization behavior. Notably, the Temperature Scaling variant alone, which requires only access to output logits, retained meaningful discriminative power, suggesting a route toward black-box approximations.

Efficiency is where PUC most clearly separates itself from the shadow-model paradigm. Benchmarked on an NVIDIA RTX 4090 against LiRA with 16 shadow models and Curv, PUC completed its full perturbation pipeline in roughly 426 seconds, a 21-fold speedup over LiRA’s approximately 9,133 seconds, and its runtime does not grow with the number of shadow models because it trains none. The trade-offs are a modest memory increase of 1.17 GiB and the need for a backward pass to compute gradients for the FGSM branch. The framework does assume a white-box threat model, requiring logit access, inference-time control of dropout, and gradient computation, but the authors argue this matches realistic scenarios including open-source model hubs, physically extractable weights on edge and IoT devices, and compromised participants in federated learning.

Defenses blunt the attack but do not eliminate it. Under label smoothing, PUC remained above an AUC of 0.8 at typical smoothing factors up to 0.1, degrading to 0.54 only at the extreme value of 0.8, a setting that would itself compromise model utility. Differential privacy via DP-SGD weakens the FGSM branch through gradient clipping and noise injection, though the Temperature Scaling and MC Dropout branches retain partial power, and a fuller evaluation of stronger defenses remains future work. The broader message for the field is that privacy risk assessment may need to move from snapshot metrics to functional probing: the way a model’s confidence bends, flattens, and decays under perturbation is itself a readable record of what it was trained on, and frameworks like PUC make that record both cheaper and sharper to decode.

Subject of Research: Membership inference attacks using predictive uncertainty curves to detect training data memorization in machine learning models

Article Title: A predictive uncertainty curve framework for membership inference attacks

Article References: Liu, S., He, Y., Zhou, Z., & Zhang, Y. (2026). A predictive uncertainty curve framework for membership inference attacks. Cybersecurity, 9(1), Article 235. https://doi.org/10.1186/s42400-026-00665-5

Image Credits: AI Generated

DOI: 10.1186/s42400-026-00665-5

Keywords: membership inference attack, predictive uncertainty, machine learning privacy, Temperature Scaling, MC Dropout, FGSM, shadow models, LiRA, ResNet-18, CIFAR-10, PathMNIST, model memorization

The study, published in the Journal of Micropalaeontology, is, according to its authors, the first attempt to apply deep learning to the automated identification of Devonian miospores. The team, led by Vitalina Lokteva and corresponding author Vahram Serobyan of the Institute of Geological Sciences of the National Academy of Sciences of Armenia, together with Martin Tetard, Pierre Breuer, and Taniel Danelian, focused on three species of enormous biostratigraphic value: Teichertospora torquata, Geminospora lemurata, and Samarisporites triangulatus. These spores are widely used to define global biozones across the ancient continents of Gondwana and Laurussia, and they play a particularly important role in hydrocarbon exploration, notably in Saudi Arabia, where Devonian spore-based stratigraphy helps geologists date and correlate the rock layers that host oil and gas reservoirs.

The material came from the Ertych section in central Armenia, one of the region’s best-known localities for palynological research. Fourteen beds of upper Frasnian age were sampled during several field campaigns, and productive miospore assemblages were recovered from twelve of them. Standard palynological preparation yielded three to five slides per sample, which were examined with a Leica DM2700 transmitted-light microscope. Every specimen that a specialist qualitatively identified as one of the three target species was photographed individually with a Flexacam C5 camera, producing a library of 209 microphotographs that formed the raw material for the machine learning pipeline.

Each of those 209 images was then annotated by hand using the web-based platform Roboflow. An operator drew a rectangular bounding box around every individual spore and recorded the image identifier, class label, and microscope magnification as metadata. The resulting detection dataset contained five classes: the three target species, which together accounted for roughly 57 percent of the annotations, an “other” class comprising non-target miospore taxa, and a background class dominated by images without miospores, filled largely with fragments of plant debris known as phytoclasts. The images were divided into training, validation, and test subsets in a 70/20/10 split, with the test set kept strictly separate so that no image used in evaluation had ever been seen during training.

The architecture of the workflow reflects a growing consensus in automated microscopy: detection first, classification second. In the detection stage, the team deployed YOLOv11, a one-stage object detection network that predicts bounding-box coordinates, objectness scores, and class probabilities in a single forward pass. Initialized with weights pretrained on the COCO image dataset and trained with a single class label of “miospore”, the detector was applied to the test micrographs using a confidence threshold of 0.7. Its performance was striking. Precision, the fraction of predicted boxes that were correct, reached 0.98; recall, the fraction of true objects the model actually found, was a perfect 1.0; and the mean average precision at an intersection-over-union threshold of 0.5 came in at 0.995, averaging 0.906 across the stricter range of thresholds from 0.5 to 0.95.

Crucially, the detection stage did more than simply find the spores. For each detection, the pipeline stored the bounding-box coordinates, a cropped image patch stripped of its cluttered background, and the area of the crop, measured in pixels and calibrated to square micrometers. This inference step produced 159 background-free spore crops, which became the classification dataset, split 70/15/15 into training, validation, and test subsets. The bounding-box area was retained as an auxiliary numeric feature, a simple piece of morphometric information that would later prove surprisingly valuable when taxa differed mainly in size.

For the classification stage, the researchers compared three convolutional neural networks, all implemented in PyTorch and trained through transfer learning, in which the pretrained convolutional feature extractors were frozen and only the task-specific heads were fine-tuned on the miospore data. VGG16, a 16-layer network that stacks small 3×3 convolutions, served as a stable, well-understood baseline. ResNet-18, an 18-layer network with residual skip connections that prevent optimization problems in deeper models, offered a lightweight and efficient feature extractor. EfficientNet-B0, the smallest member of the EfficientNet family, provided a compact architecture suited to limited data and compute. During training, images were resized to 224 by 224 pixels and augmented on the fly with flips, rotations of up to 15 degrees, color jitter, normalization, and random erasing, while validation images underwent only resizing and normalization to keep the evaluation unbiased.

The results, averaged over ten independent training runs, showed a clear hierarchy. ResNet-18 achieved the highest mean test accuracy at 87.9 percent, along with a Matthews correlation coefficient of 0.85, a metric that condenses the entire multiclass confusion matrix into a single value between minus one and one and is robust to the class imbalance typical of palynological datasets. VGG16 posted the highest macro-F1 score of 87.3 percent with a comparable test accuracy of 87.5 percent, while EfficientNet-B0 trailed on every metric, with 79.6 percent accuracy, 76.4 percent macro-F1, and an MCC of 0.75. In the best single runs, both ResNet-18 and VGG16 reached a test accuracy of 95.8 percent. The confusion matrices revealed that all three classifiers identified T. torquata flawlessly, with an F1 score of 1.00, likely because that species is large, reaching 85 to 140 micrometers, and bears a distinctive cavate structure, sub-triangular outline, and clear trilete mark. G. lemurata, a cavate spore of roughly 40 to 50 micrometers with a rounded outline, scored between 0.88 and 0.92. S. triangulatus proved hardest, with F1 scores of 0.73 to 0.80, probably because poorer preservation in the Armenian material obscures its diagnostic cone-shaped sculpture and indistinct trilete mark.

The authors are candid about the limits of the proof of concept. The classification dataset of 159 crops is small by deep learning standards, and the heterogeneous “other” class, which lumps together many different taxa, inevitably adds within-class variability. All images were acquired with a single microscope and a single preparation protocol, so performance may shift when the method meets material prepared and imaged under different conditions. Even so, the pipeline’s average per-class accuracies of 84.5 to 86.25 percent fall comfortably within the range of 83.3 to 99.8 percent reported in recent palynological and microfossil classification studies, a notable achievement for fossil material, which is consistently harder to classify than the pristine modern pollen on which most earlier models were trained. Prior work has shown that accuracy drops when models trained on modern pollen are applied to damaged or fossil specimens, making the team’s direct focus on fossil Devonian spores a meaningful step forward.

The researchers are careful to position the tool as an assistant rather than a replacement for expert knowledge. Supervised AI classification remains dependent on expert-defined taxonomic concepts and carefully curated training data, and the workflow is best understood as a screening and decision-support system that standardizes routine identifications, reduces observer bias, and improves reproducibility. The path forward is clear: enlarging the image dataset, broadening taxonomic coverage, incorporating material from additional sections and laboratories, and coupling the detector with automated slide-scanning systems to raise throughput. The code and annotated dataset have been released on Zenodo, inviting the community to build on the work. If the approach scales as hoped, the same two-step logic could extend to other fossil groups, including Paleozoic marine phytoplankton such as acritarchs and prasinophytes, and even scolecodonts, bringing the routine dating of sedimentary rocks, from academic biostratigraphy to oil exploration, into the age of machine perception.

Subject of Research: Automated detection and identification of Devonian miospores using convolutional neural networks

Article Title: Artificial intelligence applied to the automated detection and identification of Devonian miospores

Article References: Lokteva, V., Serobyan, V., Tetard, M., Breuer, P., & Danelian, T. (2026). Artificial intelligence applied to the automated detection and identification of Devonian miospores. Journal of Micropalaeontology, 45(1), 405-413. https://doi.org/10.5194/jm-45-405-2026

Image Credits: AI Generated

DOI: 10.5194/jm-45-405-2026

Keywords: artificial intelligence, convolutional neural networks, Devonian miospores, palynology, biostratigraphy, YOLOv11, ResNet-18, microfossils, upper Frasnian, Armenia, fossil spores, automated microscopy

The research, led by Nianbo Xu, Andy Chun Hang Chen, and colleagues at the University of Hong Kong and collaborating institutions, tackles a long-standing problem in reproductive medicine. For decades, pathologists have assessed endometrial receptivity by examining hematoxylin and eosin–stained biopsy specimens under the microscope, dating the tissue according to morphological criteria that describe how the lining changes across the menstrual cycle. The approach, however, is notoriously subjective. Two experienced observers can assign different dates to the same slide, and even the same observer may disagree with themselves on different days. This observer variability has undermined confidence in histological dating precisely at a moment when objective, reproducible assessment of endometrial receptivity matters most—in deciding how to time embryo transfer and counsel patients about their chances of success.

The team’s strategy was to ask whether deep learning could extract predictive information from endometrial histology that human eyes cannot reliably quantify. They built and compared two fundamentally different computational pipelines. The first was an end-to-end ResNet-18, a convolutional neural network trained from scratch on whole-slide images of endometrial biopsies, learning its own visual features directly from the data. The second leveraged UNI2-h, a so-called foundation model for pathology—a massive neural network pretrained on enormous collections of histology images spanning many organs and diseases. Rather than retraining this heavyweight model, the researchers froze its weights and used it as a fixed feature extractor, converting each slide into a rich numerical representation that a smaller classifier could then interpret. This distinction proved to be the intellectual heart of the study: can a general-purpose vision model, trained on pathology writ large, be repurposed for the specialized task of fertility assessment?

To answer that question, the researchers assembled natural-cycle endometrial biopsies and developed ensembles of models through ten-fold cross-validation, a rigorous resampling scheme in which the data are repeatedly partitioned so that every case serves as both training material and test material across different iterations. The resulting ensembles were then evaluated in an internal held-out cohort of patients with known live birth outcomes, providing an outcome-based test of whether the models’ predictions translated into clinically meaningful endpoints. Crucially, the team also designed an external, phase-based evaluation that asked a different question: could the models distinguish tissue sampled at LH+7—the mid-luteal window of maximal receptivity, seven days after the luteinizing hormone surge—from tissue sampled at other phases of the cycle? This dual evaluation exposed a striking asymmetry in how the two architectures behaved.

The whole-slide ResNet-18, trained end to end, performed well on the internal outcome-based testing but collapsed to near-random performance when asked to identify the LH+7 window in external testing. In other words, the network had apparently learned features that correlated with pregnancy outcomes in its own dataset but did not correspond to genuine, transferable knowledge of endometrial physiology. The UNI2-h-based pipeline, by contrast, showed far less divergence between the two tasks. Using mean pooling to aggregate the foundation model’s per-patch features and a multilayer perceptron as the classifier, the ensembled UNI2-h models achieved an area under the receiver operating characteristic curve—a standard measure of discriminative ability—of 0.74 plus or minus 0.03 in internal outcome-based evaluation, and a remarkably strong 0.90 plus or minus 0.05 in external phase-based testing. The foundation model, it seemed, carried visual vocabulary broad enough to recognize biologically real patterns of endometrial maturation rather than dataset-specific shortcuts.

The researchers then introduced a biologically motivated refinement. Because implantation occurs at the luminal epithelium—the single layer of cells lining the uterine cavity that makes first contact with the embryo—they built luminal epithelium–focused models that concentrated the network’s attention on this maternal-embryo interface rather than on the entire slide. The intuition is elegant: if the molecular and morphological dialogue between mother and embryo is what determines receptivity, then a model forced to look at the interface might learn more relevant signals with less data. Despite training on a smaller set of images, the LE-focused models retained performance comparable to their whole-slide counterparts. In internal outcome-based testing, LE-focused ResNet-18 and UNI2-h models achieved AUROCs of 0.71 plus or minus 0.03 and 0.74 plus or minus 0.09, respectively, while external phase-based testing yielded AUROCs of 0.80 plus or minus 0.16 for ResNet-18 and 0.84 plus or minus 0.06 for UNI2-h. The modest widening of confidence intervals in some LE-focused results reflects the reduced training volume, yet the fact that performance held up at all argues that the interface itself carries substantial predictive information.

Interpretability analysis reinforced this biological reading. Using Grad-CAM, a technique that visualizes which image regions most influence a convolutional network’s output, the researchers reviewed the LE-focused ResNet-18 models and found that attention commonly fell on the luminal epithelium alone or on the epithelium mixed with adjacent stromal tissue. The networks were, in effect, independently rediscovering the anatomical site where implantation unfolds—a reassuring sign that the models were not latching onto artifacts such as staining gradients or slide edges, but on structures with genuine physiological relevance to fertility.

Perhaps the most clinically consequential experiments came from multimodal modeling. Endometrial histology does not exist in isolation; clinicians routinely weigh hormone measurements, maternal age, endometrial thickness measured by ultrasound, and body mass index when predicting IVF success. The team therefore built integrated models that combined histology-derived features with these clinical metadata and compared them against a metadata-only baseline. The results were unambiguous: models incorporating histology significantly outperformed the clinical metadata-only model, demonstrating that the tissue slide contains predictive signal beyond what conventional clinical variables capture. An examination of the integrated model’s feature weighting revealed a hierarchy of influence—histology outputs dominated, estradiol levels and maternal age contributed smaller but appreciable weights, while progesterone, endometrial thickness, and BMI contributed negligibly. That endometrial thickness, a mainstay of clinical assessment, added essentially nothing once histology was available is a provocative finding that will likely fuel debate about which measurements deserve their place in routine IVF workups.

The implications extend beyond the IVF clinic. The study offers a case study in how foundation models—large, general-purpose networks pretrained on broad data—are changing the economics of medical AI. Where a specialized task like endometrial assessment once demanded large annotated datasets that few centers possess, repurposing a frozen foundation model allowed the Hong Kong team to achieve robust performance with a comparatively modest collection of biopsies. The contrast between the end-to-end ResNet-18, which overfit to its training distribution, and the foundation-model pipeline, which generalized across evaluation tasks, suggests that transferable visual representations are now a practical asset rather than a theoretical promise. For reproductive medicine specifically, the findings support histology-based AI as a foundation for fertility-oriented endometrial assessment, potentially offering clinicians an objective, reproducible readout of receptivity to complement or eventually replace subjective dating.

Caution is still warranted before such models reach the bedside. The internal outcome-based AUROC of roughly 0.74 indicates meaningful but imperfect discrimination, and the external phase-based results, while impressive, test a different task than live birth prediction itself. Prospective validation in independent, multi-center cohorts—and demonstration that model-guided decisions actually improve outcomes—remains the necessary next step. Yet the direction of travel is clear. A routine biopsy, read by an algorithm that sees what pathologists cannot, may soon tell a patient more honestly whether her uterus is ready to receive an embryo. In a field where hope is measured in cycles and years, an objective window into endometrial receptivity would be nothing short of transformative.

Subject of Research: Deep learning of endometrial histology to predict cumulative live birth in IVF cycles

Article Title: Foundation model-powered deep learning of endometrial histology for predicting the cumulative live birth of an in vitro fertilization cycle

Article References: Xu, N., Chen, A. C. H., Ruan, H., Yang, D., Liao, R., Qi, X., Fong, S. W., Cao, D., Yu, L., Huang, Y., Yeung, W. S. B., Ng, E. H. Y., & Lee, Y. L. (2026). Foundation model-powered deep learning of endometrial histology for predicting the cumulative live birth of an in vitro fertilization cycle. PLOS Digital Health, 5(9), e0001744. https://doi.org/10.1371/journal.pdig.0001744

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001744

Keywords: in vitro fertilization, endometrial receptivity, deep learning, foundation models, histopathology, cumulative live birth, luminal epithelium, Grad-CAM, multimodal prediction, reproductive medicine, PLOS Digital Health, ResNet-18

Neonatal pneumothorax is a deceptively difficult diagnostic target. It occurs in roughly 0.05 to 2 percent of all newborns and in as many as 9.2 percent of very preterm infants who require respiratory support. While many cases resolve on their own, others deteriorate rapidly into hypoxemia or hemodynamic instability demanding urgent intervention. The standard screening tool, a supine anteroposterior chest X-ray, is far from perfect. A recent meta-analysis reported pooled sensitivity of 82 percent and specificity of 96 percent for physician-interpreted supine neonatal chest radiography. The fundamental problem is geometric: when a baby lies on its back, free air collects anteriorly or medially in the chest rather than at the lung apex, exactly where radiologists are trained to look in upright adults. Combine that with a tiny thoracic volume, mediastinal overlap, poor inspiration, and often coexisting lung disease, and even skilled readers can overlook a clinically significant air leak.

Existing artificial intelligence tools do not simply transfer to this population. When software approved for adults was applied to pediatric chest radiographs, diagnostic performance varied sharply by age, and children aged two years or younger accounted for 81.5 percent of the incorrect predictions. That sobering statistic motivated the research team to build a model designed specifically for the youngest patients. Their approach relied on a ResNet-18 convolutional neural network, initialized with ImageNet weights and then further pre-trained in a self-supervised manner on 1,802 publicly available normal adult chest radiographs, allowing the encoder to learn general radiographic features such as lung-field structure, rib and soft-tissue contrast, and texture before ever encountering a neonatal image.

The training pipeline was architecturally sophisticated. After self-supervised pre-training, the researchers fine-tuned a teacher model on neonatal radiographs, then trained a student model of identical architecture using knowledge distillation, a technique in which the student learns not only from hard binary labels but also from the softened probability distributions produced by the teacher. The team reports that the full configuration, combining adult-chest pre-training with distillation, outperformed both plain ImageNet initialization and adult pre-training without distillation on the test set. Each radiograph was preprocessed into a three-channel image containing original, lighter, and darker versions, a clever trick to simulate the brightness variability of real bedside imaging and force the model to learn exposure-insensitive features.

The dataset itself was formidable. Drawing on fourteen years of records, from January 2011 to December 2024, at a single tertiary university hospital, the researchers assembled 648 pneumothorax radiographs from 288 neonates and 5,511 control radiographs from 3,377 neonates who never developed the condition. A pediatric surgeon and a neonatologist, each with more than a decade of clinical experience, jointly reviewed every image to establish reference labels, recording laterality and anatomical distribution of the air collections. Data were split at the patient level into training, validation, and test sets in a 7:1:2 ratio, ensuring that no infant’s images leaked across partitions. An eight-fold cross-validation procedure then fixed the decision threshold at 0.06, selected to maximize specificity while keeping sensitivity at or above 90 percent, a prespecified constraint reflecting the model’s intended role as a screening and decision-support tool where missed cases are the costliest errors.

On the held-out test set of 1,231 images, the model delivered an AUC of 0.975, with sensitivity of 87.6 percent, specificity of 95.3 percent, accuracy of 94.5 percent, and a negative predictive value of 98.5 percent, meaning a negative result was highly trustworthy. Patient-level analyses, whether using the earliest radiograph per neonate or averaging probabilities across all of a baby’s films, yielded nearly identical AUCs of 0.976 and 0.977. The Brier score of 0.0309 indicated reasonable calibration, although the calibration curve showed some deviation in the intermediate and high probability ranges where few observations existed. The positive predictive value of 68.5 percent, however, means roughly one in three positive alerts was a false positive, which the authors caution could burden clinicians with unnecessary urgent reviews if the model were deployed naively as a standalone rule-in test.

Interpretability analysis added an intriguing wrinkle. Using gradient-weighted class activation mapping, or Grad-CAM, the team examined whether the model’s visual attention landed on the correct lung. Among 163 lung-level evaluations, activation matched the clinically determined side in 91.3 percent of right-lung instances but only 67.6 percent of left-lung instances, a statistically significant disparity. The researchers attribute this asymmetry partly to the greater radiographic complexity of the left hemithorax, where overlap from the cardiac silhouette and mediastinal structures obscures pleural air, and partly to dataset imbalance: right-sided pneumothoraces appeared in 80.7 percent of the positive films versus 64.8 percent for the left side, leaving the model with fewer examples of left-sided patterns, particularly lateral lesions, to learn from.

Perhaps the most clinically revealing finding came from logistic regression of misclassification factors. Transient tachypnea of the newborn, a common condition marked by perihilar congestion and delayed clearance of fetal lung fluid, was independently associated with model error, with an odds ratio of 1.88, while gestational age also contributed modestly with an odds ratio of 1.10 per week. Transient tachypnea appeared in 32.7 percent of false positives and fully half of all false negatives. The interpretation is that deep networks read the whole image rather than isolated signs, so background pulmonary abnormalities can either mimic pneumothorax patterns or camouflage true ones. The model, in other words, is sensitive not just to the pathology itself but to the radiographic context in which it sits, a lesson likely to generalize across pediatric AI systems.

The authors are careful about scope. The study is retrospective and single-center, the reference standard rested on consensus radiograph review rather than ultrasound or CT confirmation, no pediatric radiologist participated in annotation, and no direct comparison with physician readers was performed. External and prospective validation is explicitly warranted before any clinical implementation. Yet the trajectory is clear and exciting. The performance achieved here is comparable to or better than adult pneumothorax AI models, which typically report AUCs between 0.87 and 0.98, and it slightly exceeds NeoCLIP, a recent self-supervised neonatal foundation model that achieved an AUC of 0.93 for pneumothorax across multiple findings. Positioned as a screening adjunct rather than an autonomous diagnostician, with positive outputs always paired with clinical judgment, this neonate-specific model suggests that age-tailored AI can finally give the smallest and most vulnerable patients the diagnostic safety net their adult counterparts have begun to enjoy.

Subject of Research: Development of a deep learning model for detecting pneumothorax on supine chest radiographs of newborns

Article Title: Development and evaluation of a deep learning model for computer-aided diagnosis of neonatal pneumothorax on chest radiographs

Article References: Sanmoto, Y., Gao, Q., Hitaka, D., Zhu, X., & Masumoto, K. (2026). Development and evaluation of a deep learning model for computer-aided diagnosis of neonatal pneumothorax on chest radiographs. Pediatric Radiology. https://doi.org/10.1007/s00247-026-06778-w

Image Credits: AI Generated

DOI: 10.1007/s00247-026-06778-w

Keywords: artificial intelligence, deep learning, neonatal pneumothorax, chest radiography, neonatology, ResNet-18, knowledge distillation, Grad-CAM, computer-aided diagnosis, pediatric radiology, transient tachypnea of the newborn, NICU