The core insight behind the work is that existing reconstruction-based anomaly detectors tend to sacrifice one kind of fidelity for another. When a model is trained to reproduce normal images, it must capture both the global semantic structure of a scene and the fine local texture of its surfaces. Networks that excel at global consistency often produce blurred reconstructions in which subtle defects are smoothed away, while networks tuned for local detail can lose track of the overall arrangement of objects and misinterpret normal variation as abnormal. This trade-off between semantic coherence and texture fidelity has been a persistent bottleneck, and it directly limits how precisely a system can localize an anomaly down to individual pixels.

The Residual Gated AutoEncoder addresses the global side of this problem. Built on top of features extracted by DINOv2, a powerful self-supervised vision transformer pretrained on massive image collections, the autoencoder learns to reconstruct the semantic essence of normal images. Its distinguishing feature is a gating mechanism that controls how information flows through residual connections inside the network. Rather than letting every feature propagate indiscriminately, the gates suppress redundant signals while preserving fine details, which keeps the reconstruction sharp where it matters. The result is a global reconstruction pathway that maintains the identity and layout of objects without washing out the small cues that often betray a defect.

In parallel, the Patchwise Nearest Neighbor Reconstruction module takes a deliberately different route. Instead of learning to generate an image, it retrieves. The researchers first build a memory bank of local feature patches extracted exclusively from normal training samples. When a new image arrives, PNNR reconstructs each local patch by finding its nearest neighbors in this memory bank and assembling a reconstruction from the retrieved pieces. Because the memory bank contains only normal material, any patch that cannot be matched well, for instance a scratch on a metal surface or a crack in a circuit board, will be reconstructed poorly. This retrieval-based strategy preserves fine-grained local structure with a fidelity that learned generators struggle to match.

The final anomaly map emerges from the disagreement between the two pathways. The system computes pixel-wise discrepancies between the output of the Residual Gated AutoEncoder and the output of the PNNR module, producing a spatial map of where the two reconstructions diverge. Because the two modules fail in complementary ways, their disagreement is informative: regions that look anomalous to one pathway but not the other, or to both, can be flagged with high confidence. The authors emphasize that this comparison yields interpretable localization, meaning a human inspector can see exactly which pixels drove the decision rather than receiving an opaque score for the whole image.

The entire framework rests on DINOv2 features, and that choice matters. DINOv2 was trained without labels using self-supervised objectives, and it produces visual representations that transfer remarkably well across domains. By anchoring both reconstruction pathways to these pretrained features, RGAE+PNNR avoids training a full vision backbone from scratch, which reduces data requirements and helps the system generalize. The authors report that the architecture generalizes effectively across diverse domains, from the complex real-world industrial scenes of the VisA dataset to the controlled object surfaces of MVTec AD, two benchmarks that stress very different aspects of visual appearance.

Extensive experiments on both benchmarks showed that the proposed architecture achieves improved image-level and pixel-level AUROC scores compared to state-of-the-art methods. The area under the receiver operating characteristic curve is the standard yardstick in this field: the image-level metric measures whether a whole picture is correctly judged normal or defective, while the pixel-level metric measures whether the defect is localized accurately within it. Achieving gains on both simultaneously is notable, because methods that boost whole-image classification often do so at the expense of precise segmentation, and vice versa. The dual-path design appears to sidestep that compromise by letting each pathway specialize in the regime where it is strongest.

The study situates itself within a rapidly evolving landscape. Earlier approaches such as PaDiM modeled the statistical distribution of patch features, while PatchCore pushed memory-bank nearest-neighbor ideas toward total recall on industrial data. More recent systems have explored normalizing flows with FastFlow, efficient millisecond-level inference with EfficientAD, and transformer-based designs such as Dinomaly, which applies a less-is-more philosophy to multi-class detection. Other lines of work have incorporated vision-language models like CLIP for zero-shot detection, and dual memory banks for real-world conditions. RGAE+PNNR draws on the memory-bank tradition through its PNNR module but pairs it with a learned global reconstructor, a combination the authors argue resolves the semantic-versus-texture tension more cleanly than either strategy alone.

The practical implications extend well beyond leaderboard numbers. In industrial inspection, unsupervised detectors allow manufacturers to spot defects on production lines without curating large defect libraries for every product, a requirement that has historically made automated quality control brittle and costly. In medical imaging, where pathological examples are rare and privacy constraints limit data sharing, a system that learns only from normal anatomy could flag tumors, lesions, or tissue irregularities without ever being told what they look like. In autonomous perception, vehicles encounter novel hazards constantly, and a detector that notices when the world deviates from its learned notion of normal could serve as an early-warning layer. The interpretable anomaly maps produced by this framework are especially valuable in these settings, because operators need to trust and verify machine judgments before acting on them.

The researchers have released their code publicly on GitHub, and both evaluation datasets, MVTec AD and VisA, are freely available, which lowers the barrier for other teams to reproduce, scrutinize, and extend the results. The work, authored by Nighil Natarajan, Nithilan M, Raghav Sridharan, and Chandrakala S, who contributed equally, reflects a broader trend in machine intelligence: rather than building ever-larger monolithic models, researchers are increasingly composing specialized modules whose complementary strengths can be exploited systematically. By teaching one network to remember what normal looks like globally and another to retrieve what normal looks like locally, and then listening to where the two disagree, the Chennai team offers a template that could shape anomaly detection systems well beyond the factory floor. As machines are asked to watch over more of the physical world, methods that make their judgments both accurate and explainable will only grow in importance.

Subject of Research: Unsupervised visual anomaly detection using a dual-path reconstruction architecture built on DINOv2 features

Article Title: Residual gated AutoEncoder with patchwise nearest neighbor reconstruction for visual anomaly detection

Article References: Natarajan, N., M, N., Sridharan, R., & S, C. (2026). Residual gated AutoEncoder with patchwise nearest neighbor reconstruction for visual anomaly detection. Applied Intelligence, 56(15), Article 484. https://doi.org/10.1007/s10489-026-07530-5

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07530-5

Keywords: visual anomaly detection, unsupervised learning, autoencoder, DINOv2, nearest-neighbor reconstruction, computer vision, industrial inspection, MVTec AD, VisA, AUROC, anomaly localization, self-supervised learning

The research, led by Eva Reisig, Lie Cai, Maria Müller, Benedikt Schäfgen and André Pfob of the Breast Unit at Heidelberg’s Department of Obstetrics and Gynecology, focused on a milestone known as pathological complete response, or pCR. When breast cancer patients receive neoadjuvant treatment — chemotherapy or other systemic therapy given before surgery rather than after — pathologists examine the tissue removed during the operation. If they find no residual invasive cancer in the breast or the sampled lymph nodes, the patient is said to have achieved a complete pathological response. This outcome is more than a reassuring label: it is one of the strongest known predictors of long-term survival and freedom from recurrence, and it has become a key endpoint in breast cancer drug trials.

The clinical logic behind the study is straightforward but ambitious. If a model could reliably predict, before surgery, that a patient’s tumor has been completely eradicated, that information could spare her an operation that would remove nothing but scar tissue. Breast-conserving therapy and even sentinel lymph node procedures carry physical and psychological costs, and for patients whose disease has vanished, those costs may buy little benefit. The obstacle has always been accuracy. Clinical examination, ultrasound, mammography and magnetic resonance imaging each offer imperfect glimpses of the tumor’s status, and no single modality is trustworthy enough on its own to justify withholding surgery. The Heidelberg team’s wager was that combining all of these signals — along with patient and tumor characteristics — in a single algorithmic framework might push predictive performance into clinically useful territory.

To build that framework, the researchers assembled a retrospective, single-center cohort of 388 cases of invasive breast cancer treated with neoadjuvant therapy. For each case, they extracted an unusually rich set of 111 variables spanning three imaging modalities — ultrasound, mammography and magnetic resonance imaging — together with histopathological and clinical information such as receptor status and patient demographics. This breadth matters. Much prior work on predicting pCR has leaned on a single imaging technique or a narrow panel of biomarkers, leaving the algorithm blind to complementary information. By pooling findings from the full diagnostic workup that every breast cancer patient already undergoes, the team aimed to approximate the way an experienced multidisciplinary tumor board weighs evidence, but with mathematical consistency.

Raw data alone, however, do not make a model. With 111 candidate variables and fewer than 400 cases, the researchers faced the classic statistical trap of overfitting — the risk that an algorithm learns quirks of its training data rather than genuine biological relationships. Their first line of defense was a Pearson Product-Moment correlation analysis, which identified variables that were essentially measuring the same underlying quantity. Eight features were removed on these grounds, trimming redundancy before any predictive modeling began. The second filter was LASSO regularization, a technique from statistical learning that shrinks the coefficients of weak predictors toward zero and effectively eliminates them, keeping only the variables that earn their place through genuine association with the outcome. After this two-stage selection, 14 variables survived as the final predictors.

The modeling choice itself was deliberately conservative. Rather than deploying a black-box deep learning architecture, the team used logistic regression with an elastic net penalty, abbreviated GLM in the paper. Elastic net regularization blends the variable-selection behavior of LASSO with the ability to handle correlated predictors, making it well suited to clinical datasets where imaging findings and receptor statuses often travel together. Logistic regression also produces transparent, interpretable coefficients — each variable’s contribution to the prediction can be inspected and questioned, a property that matters enormously when the output will inform decisions about whether a patient undergoes surgery. The authors note that this combination of LASSO-based selection and regularized regression yields interesting insights into the predictive potential of individual variables, all of which are available as part of standard clinical practice, with no experimental tests required.

The results offer cautious encouragement. In the total cohort, pathological complete response was confirmed in 158 of 388 cases, a pCR rate of 40.7 percent — a reminder of how common complete tumor eradication already is under modern neoadjuvant regimens. When the final model was evaluated on a validation set of data it had not learned from, it achieved an area under the receiver operating characteristic curve, or AUROC, of 0.76. In practical terms, an AUROC of 0.5 reflects coin-flip performance while 1.0 represents perfect discrimination; 0.76 indicates the model separates responders from non-responders meaningfully better than chance, though not yet with the near-certainty that would allow surgeons to abandon operative assessment outright.

Two secondary metrics sharpen the clinical picture. The model’s false positive rate — the proportion of patients without pCR who were incorrectly flagged as complete responders — was 8.7 percent, a figure the authors highlight because false positives are the dangerous errors in this setting: a patient wrongly predicted to have no residual tumor might be denied a surgery she actually needed. Meanwhile, the positive predictive value reached 75 percent, meaning that when the model declared a complete response, it was right three times out of four. For context, clinical complete response assessed by conventional means has historically lagged well behind pathological confirmation, which is precisely why surgery has remained non-negotiable. A model that confines its confident predictions to cases where it is right three-quarters of the time, while keeping false alarms below nine percent, represents a measurable step toward closing that gap.

The study’s limitations are the natural constraints of its design. It was retrospective and monocentric, drawing on data from a single German institution, which means the model must be validated externally before any change to surgical practice could be contemplated. The pCR rate in the cohort, while within the range reported for modern neoadjuvant therapy, reflects the specific patient mix and treatment protocols of one center. And even a well-calibrated prediction of complete response would need to be weighed against the possibility of residual disease in areas that imaging cannot fully interrogate. The authors are careful to frame their algorithm as a promising approach to improving prediction, not as a replacement for pathological assessment.

Even so, the significance of the work lies less in its headline numbers than in its method and its message. It demonstrates that a parsimonious, interpretable model built exclusively from routinely available data — three imaging modalities plus standard clinicopathological variables — can approach the problem of treatment response prediction with respectable accuracy, without exotic inputs or opaque architectures. As neoadjuvant therapy becomes more effective and more widely used, the population of patients whose tumors vanish before surgery will only grow, and the question of who truly needs an operation will press harder on the field. Studies like this one sketch the statistical scaffolding for that conversation, and they suggest that the path to personalized, response-adapted breast cancer care may run not through futuristic diagnostics, but through smarter use of the data clinicians already hold in their hands.

Subject of Research: Machine learning prediction of pathological complete response after neoadjuvant chemotherapy in breast cancer

Article Title: Machine learning model for the prediction of pathological complete response after neoadjuvant chemotherapy in breast cancer

Article References: Reisig, E., Cai, L., Müller, M., Schäfgen, B., & Pfob, A. (2026). Machine learning model for the prediction of pathological complete response after neoadjuvant chemotherapy in breast cancer. BMC Cancer. https://doi.org/10.1186/s12885-026-17115-7

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17115-7

Keywords: breast cancer, machine learning, neoadjuvant chemotherapy, pathological complete response, predictive medicine, LASSO regularization, logistic regression, MRI, ultrasound, mammography, AUROC, surgery

The study, led by Hao Pang, Fiseha Berhanu Tesema, Tianxiang Cui, Yuan Cheng, and Yanwen Mao, introduces a model called BCAG-DTI, short for Bidirectional Cross-Attention and Global Aggregation drug-target interaction model. Rather than simply claiming superiority over the widely used MolTrans baseline, the researchers ran eight carefully controlled configurations that separately toggled bidirectional cross-attention, a global average-and-max pooling layer, the design of the classification head, and the optimisation strategy. By changing one factor at a time across fixed training, validation, and test partitions with five matched random seeds, they could attribute each gain to its true source, a level of experimental hygiene that remains rare in a literature crowded with simultaneous architectural and training tweaks.

The headline results are striking. On the BindingDB benchmark, the complete BCAG-DTI configuration lifted the mean area under the receiver operating characteristic curve, or AUROC, from 0.8815 for MolTrans to 0.9063. On BIOSNAP, the corresponding figure rose from 0.8631 to 0.8891, and on DAVIS from 0.8808 to 0.8955. Because AUROC alone can flatter models on imbalanced datasets, the team also measured the area under the precision-recall curve, AUPRC, which is often the more honest metric for interaction prediction. There the improvements were even more pronounced: gains of 0.0831 on BindingDB, 0.0346 on BIOSNAP, and 0.0619 on DAVIS, representing substantial advances in the model’s ability to rank true interactions above false ones.

The most provocative finding, however, lies in the ablation analysis that followed. When each component was tested in isolation, the enhanced classifier design emerged as the single most influential element, achieving the highest mean AUROC on BindingDB and BIOSNAP and the highest mean AUPRC and F1-score on all three datasets. The enhanced training strategy also delivered substantial improvements on its own. By contrast, bidirectional cross-attention alone actually reduced performance when paired with the baseline training configuration. In other words, the glamorous cross-modal attention machinery that gives the model its name contributes value only under the right optimisation conditions, and its benefit is dataset-dependent rather than universal. On DAVIS, it was the combination of cross-attention, global pooling, and enhanced training that achieved the highest mean AUROC, underscoring how sensitive these components are to context.

This pattern carries a lesson that extends well beyond one model family. Transformer-style cross-attention between molecular and protein representations has become a fashionable design choice in bioinformatics, frequently credited with allowing models to learn which parts of a ligand speak to which residues of a binding pocket. The Nottingham results suggest that such credit may sometimes be misplaced, or at least overstated, because classifier head design and optimisation schedules can account for a large share of the reported improvement. For a field racing to publish ever-larger architectures, the message is that rigorous component-wise evaluation is not optional bookkeeping but the difference between genuine scientific progress and an artefact of training procedure.

Robustness testing added another layer of nuance. In experiments on BIOSNAP designed to simulate real-world sparsity, BCAG-DTI outperformed MolTrans when predicting interactions for unseen drugs, for unseen proteins, and in settings where 70 to 90 percent of interaction labels were missing. These scenarios matter enormously in practice, because the vast majority of possible drug-protein pairs have never been measured, and any screening tool must generalise beyond the sparse matrix of known interactions. The one blemish appeared at 95 percent missing data, where BCAG-DTI’s AUROC and F1-score dipped slightly below the baseline, a reminder that even improved models have limits when supervision becomes extremely thin.

To situate their results against the broader literature, the authors also evaluated an external reference model, CPI-GGS, on the same fixed BIOSNAP partitions. CPI-GGS achieved 0.8619 plus or minus 0.0023 AUROC, 0.8645 plus or minus 0.0042 AUPRC, and 0.7938 plus or minus 0.0028 F1-score, compared with 0.8891 plus or minus 0.0078, 0.8992 plus or minus 0.0067, and 0.8178 plus or minus 0.0094 for BCAG-DTI. Crucially, the team interpreted this comparison with appropriate caution, noting that the two systems use different input preprocessing pipelines and differ substantially in model capacity. Rather than declaring outright victory, they framed the numbers as a same-split reference point, a refreshing contrast to the practice of comparing scores across incompatible datasets and splits that still plagues the field.

Perhaps the most intellectually honest portion of the paper concerns attention interpretability. Cross-attention weights are often visualised and marketed as evidence that a model has learned genuine binding contacts between a ligand and a protein pocket. The authors’ case analysis pushes back on this convention, indicating that cross-attention weights should be treated as model-internal allocation patterns rather than validated binding contacts. This distinction matters for downstream users: a researcher who trusts an attention heatmap as a structural hypothesis could waste laboratory resources chasing a pattern that reflects the model’s internal bookkeeping rather than physical chemistry. It is a caution that resonates across the growing literature on attention-based models in the life sciences.

The study’s methodology deserves emphasis as a model for the field. Fixing the data partitions across all principal experiments and running five matched random seeds allowed the authors to report means with a meaningful sense of variance, guarding against the lucky-seed effect that can inflate single-run results. The three benchmarks themselves span different regimes: BindingDB offers a large, chemically diverse collection of measured affinities, BIOSNAP provides a network-derived interaction dataset well suited to cold-start and missing-data experiments, and DAVIS presents a smaller, kinase-focused panel. Showing consistent gains across all three, while honestly reporting where components underperform, lends the conclusions a credibility that single-benchmark papers rarely achieve.

For drug discovery pipelines, the practical implications are immediate. Teams building interaction predictors should scrutinise their classification heads and optimisation strategies before investing in architectural complexity, since the cheapest improvements may lie in the least glamorous parts of the stack. Cross-attention and pooling modules should be adopted with the understanding that their value is conditional, emerging only in combination with appropriate training regimes and on certain data distributions. And anyone publishing comparisons against established baselines should take a page from this study’s playbook: hold the data fixed, vary one thing at a time, and report what actually moved the needle. As machine learning continues to reshape how candidate drugs are prioritised for synthesis and testing, evaluations of this rigour will be essential to ensure that the field’s progress is real rather than an artefact of experimental design.

Subject of Research: Controlled ablation evaluation of architectural, classifier, and training refinements in deep learning-based drug-target interaction prediction

Article Title: Controlled evaluation of architectural, classifier, and training refinements in MolTrans-based drug-target interaction prediction

Article References: Controlled evaluation of architectural, classifier, and training refinements in MolTrans-based drug-target interaction prediction. (n.d.). https://doi.org/10.1186/s12859-026-06634-6

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06634-6

Keywords: drug-target interaction, MolTrans, cross-attention, ablation study, deep learning, BindingDB, BIOSNAP, DAVIS, AUROC, classifier design, optimisation, bioinformatics

Now, a team of researchers at the U.S. Food and Drug Administration’s Division of Imaging, Diagnostics, and Software Reliability has developed a new tool designed to catch these flaws before they can do damage. The system, called HistoART, is described in a study published in the journal Neural Computing and Applications. Led by Seyed Kahaki, with contributions from Alexander Webber, Ghada Zamzmi, Adarsh Subbaswamy, Rucha Deshpande, and Aldo Badano, the work compares three fundamentally different strategies for detecting artifacts in whole slide images, and finds that a large foundation model originally trained for general pathology tasks outperforms both a conventional deep learning network and a classical, hand-engineered approach.

The stakes are higher than they might first appear. As hospitals adopt digital pathology workflows, the reliability of downstream image analysis tasks, from tumor detection to grading, depends on the quality of the images feeding into them. An algorithm trained to recognize cancer cells may misinterpret a tissue fold as a suspicious structure, or an air bubble as a region of lost tissue. Previous studies have documented how quality control, or the lack of it, directly affects the accuracy of computational pathology systems. Yet many existing quality control tools rely on fixed rules and handcrafted measurements that may not generalize across the wide variety of scanners, staining protocols, and tissue types encountered in real-world laboratories.

HistoART tackles the problem with a three-way comparison. The first approach, the foundation model-based approach, fine-tunes UNI, a general-purpose foundation model for computational pathology that has been trained on massive collections of pathology images. Foundation models of this kind learn rich, general-purpose visual representations that can be adapted to many downstream tasks with relatively little additional training. The second approach, the deep learning approach, is built on a ResNet50 backbone, a widely used convolutional neural network architecture originally developed for general image recognition. The third, the knowledge-based approach, dispenses with learned representations altogether and instead relies on handcrafted features derived from texture, color, and frequency-based metrics, drawing on decades of classical image analysis research.

All three approaches were trained and evaluated to detect six of the most prevalent artifact types in whole slide images: tissue folds, out-of-focus regions, air bubbles, tissue damage, marker traces, and blood contamination. Each of these artifacts arises at a different stage of the workflow. Tissue folds occur when sections of tissue wrinkle during mounting on the slide. Out-of-focus regions result from imperfect autofocus during scanning. Air bubbles become trapped under the coverslip, marker traces are left by pens used to label slides, and blood contamination can obscure underlying tissue architecture. Detecting such a heterogeneous set of defects demands a system capable of recognizing very different visual patterns, from the sharp linear creases of a fold to the diffuse haze of an out-of-focus patch.

The evaluation was deliberately broad. The researchers assembled a dataset of more than 50,000 image patches sourced from diverse whole slide imaging scanners, including instruments from Hamamatsu, Philips, and Leica Aperio AT2, and drawn from multiple imaging sites. This diversity matters because artifacts can look different depending on the scanner that produced the image, and a detection system that works well on one manufacturer’s output may fail on another’s. By testing across scanners and sites, the team aimed to assess how well each approach would generalize beyond a single laboratory environment, a persistent weakness of many published machine learning studies in pathology.

The results were striking. The foundation model-based approach achieved a patch-wise area under the receiver operating characteristic curve, or AUROC, of 0.995, with a 95 percent confidence interval of 0.994 to 0.995, a level of performance that approaches the theoretical ceiling of a perfect classifier. The ResNet50-based deep learning approach reached an AUROC of 0.977, with a confidence interval of 0.977 to 0.978, while the knowledge-based approach, built on handcrafted features, trailed at 0.940, with a confidence interval of 0.933 to 0.946. The AUROC metric measures a classifier’s ability to distinguish between positive and negative cases across all possible decision thresholds, so the gap between 0.995 and 0.940 represents a substantial difference in real-world reliability, particularly for the subtle or ambiguous artifacts that matter most in clinical practice.

Why would a foundation model so decisively outperform a purpose-built convolutional network? The likely answer lies in the breadth of its pretraining. UNI was trained on enormous and varied collections of pathology imagery, meaning it has already encountered an extraordinary range of tissue appearances, staining variations, and imaging conditions before any artifact-specific fine-tuning begins. When subsequently adapted to the artifact detection task, it brings that prior knowledge to bear, allowing it to distinguish, say, a genuine tissue fold from a naturally occurring tissue edge with greater confidence than a network learning visual features from scratch. The handcrafted approach, meanwhile, is limited by the imagination of its designers: texture descriptors, color statistics, and frequency-domain measures capture certain artifact signatures well but may miss patterns that do not fit predefined mathematical descriptions.

Detection alone, however, is not enough. A laboratory needs to know what to do with the information. To bridge the gap between detection and actionable conclusions, the team developed a quality report scorecard that quantifies the number of high-quality image patches in a dataset and visualizes the distribution of artifact subgroups. This reporting layer transforms raw per-patch predictions into a summary that laboratory staff and algorithm developers can act upon: a slide riddled with folds can be rescanned or excluded, a scanner producing systematic out-of-focus regions can be serviced, and a dataset destined for training a diagnostic model can be filtered to remove compromised patches. In this way, HistoART functions not merely as a classifier but as a quality assurance pipeline for the entire digital pathology workflow.

The researchers describe the resulting system as a multi-branch pipeline that may enhance the reliability of whole slide image analysis by providing a scalable and interpretable approach for improving digital pathology workflows. Interpretability is a key selling point of the scorecard design, since regulators and clinicians are often wary of black-box systems that flag images without explaining why. By visualizing which artifact types dominate a slide or dataset, the tool offers a transparent window into the failure modes of the imaging process. Consistent with the team’s commitment to open science, all the data and methods from the work are available for download on GitHub, allowing other laboratories to adopt, scrutinize, and extend the system. As whole slide imaging continues its march toward becoming the default medium of diagnostic pathology, tools like HistoART suggest that the quality of the digital slide, long treated as an afterthought, is finally being engineered with the same rigor as the diagnoses that depend on it.

Subject of Research: Artifact detection in whole slide histopathology images using foundation models, deep learning, and handcrafted features

Article Title: HistoART: Histopathology artifact detection based on large foundation model

Article References: Kahaki, S., Webber, A., Zamzmi, G., Subbaswamy, A., Deshpande, R., & Badano, A. (2026). HistoART: Histopathology artifact detection based on large foundation model. Neural Computing and Applications, 38(18), Article 753. https://doi.org/10.1007/s00521-026-12454-9

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12454-9

Keywords: histopathology, whole slide imaging, artifact detection, foundation model, UNI, ResNet50, digital pathology, quality control, machine learning, AUROC, cancer diagnostics, image analysis

The research team, led by investigators at Children’s Hospital, Zhejiang University School of Medicine, together with collaborators spanning thirteen pediatric intensive care units across China, assembled a retrospective cohort of 2,374 children under eighteen years of age who were admitted to intensive care with suspected infections between 2023 and 2025. Rather than treating this population as a single undifferentiated group, the investigators prespecified two clinically distinct sub-cohorts for focused analysis: a neonatal group of 576 infants aged twenty-eight days or younger, and a hematology group of 494 children, a population in which underlying malignancies, immunosuppression, and hematologic abnormalities complicate both the diagnosis of sepsis and the interpretation of laboratory-based severity scores. This design choice reflects a growing recognition in the field that a scoring system validated in a general pediatric intensive care population may not transfer cleanly to the extremes of age or to immunocompromised patients.

Methodologically, the study followed a disciplined approach to score calculation. For each child, both the Phoenix sepsis score and the pSOFA score were computed using the worst available physiological and laboratory values recorded within the first twenty-four hours of intensive care unit admission. This window matters: early severity assessment is precisely what clinicians need when deciding how aggressively to resuscitate, monitor, and escalate care. The primary outcome was in-hospital mortality, and the principal metric of prognostic accuracy was the adjusted area under the receiver operating characteristic curve, or AUROC, a statistic that captures how well a score discriminates between children who died and those who survived, with 0.5 representing chance-level performance and 1.0 representing perfect discrimination. By adjusting the AUROC, the researchers accounted for confounding structure in the data, strengthening the comparability of the two scores.

The headline finding is statistically clear. Across the full cohort of 2,374 critically ill children, the Phoenix sepsis score achieved an adjusted AUROC of 0.840, significantly outperforming the pSOFA score’s 0.817, with a P value below 0.001. In absolute terms the difference may appear modest, but in the world of mortality prediction models, where incremental gains in discrimination can translate into earlier recognition and redirected resources for the sickest patients, such a difference across a cohort of this size carries weight. The result provides external validation, on a scale and in a healthcare context distinct from the original derivation studies, that the Phoenix framework’s emphasis on four organ systems, respiratory, cardiovascular, coagulation, and neurologic, captures lethal trajectories in infected children at least as well as, and slightly better than, the older pSOFA construct.

The picture grows more complicated in the predefined sub-cohorts. In the hematology group, the Phoenix score again outperformed pSOFA, with adjusted AUROCs of 0.718 versus 0.682, a difference that reached statistical significance at P equal to 0.006. Yet the investigators were careful to characterize discriminative ability in this subgroup as limited for both tools. Children with hematologic disease present a notoriously difficult prediction problem: baseline cytopenias, chemotherapy-related organ toxicity, and altered inflammatory responses can all inflate or deflate organ dysfunction scores independent of sepsis severity. A score of 0.718, while better than chance, leaves substantial room for misclassification, and the authors’ framing suggests that neither score should be treated as a reliable standalone prognostic instrument in this vulnerable population.

In neonates, the story changes again. Among the 576 infants aged twenty-eight days or younger, the Phoenix score posted a strikingly high adjusted AUROC of 0.915 compared with 0.850 for pSOFA, but the difference did not reach statistical significance, with P equal to 0.222. The absence of a significant gap does not mean either score performed poorly; indeed, both achieved excellent discrimination in this low-event-rate setting. Rather, the finding suggests that in newborns the two instruments are statistically interchangeable for mortality prediction, an important practical conclusion given that the Phoenix criteria were designed to span pediatric and neonatal populations and that neonatal sepsis has long posed unique challenges, from nonspecific clinical presentation to the scarcity of neonate-specific validation data for adult-derived scoring frameworks.

Beyond discrimination, the study examined how the scores behave at their operational threshold, the cutoff clinicians actually use to flag sepsis. At the standard threshold of two or more points, both the Phoenix score and pSOFA maintained high sensitivity, exceeding 80 percent across all cohorts, meaning they rarely missed a child who went on to die. The trade-off came in specificity: positive predictive values were low, at or below 50 percent everywhere, and in the neonatal sub-cohort they fell below 8 percent. This asymmetry is not a flaw unique to these scores but a mathematical consequence of applying screening thresholds in settings where deaths are, thankfully, uncommon. A highly sensitive tool catches nearly everyone at risk, but in a population where most children survive, most flagged children will nonetheless survive, producing many false alarms.

The trade-off between sensitivity and positive predictive value took on a different character in the hematology sub-cohort, where the event rate was high enough to make predictive values more meaningful. There, the Phoenix score traded approximately five percentage points of sensitivity for a meaningfully higher positive predictive value than pSOFA. In practical terms, the Phoenix score was somewhat more likely to miss a child who died, but when it flagged a child, that flag was more often correct. Which trade-off a clinician or unit should prefer depends on context: in settings where missing a death is the dominant concern, sensitivity may reign; where flagged patients trigger resource-intensive interventions, a higher positive predictive value reduces unnecessary escalations. The study does not prescribe a choice but equips decision-makers with the quantitative texture needed to make one.

The findings land amid an active international debate about the Phoenix criteria, which were developed and validated by a global consensus effort published in JAMA in 2024 and have since undergone external validation in cohorts from the United States, Bolivia, Australia, and beyond, with mixed results in low-resource settings and emergency departments. The new Chinese multicenter analysis adds a critical piece: evidence from a large, geographically diverse, non-Western cohort, with dedicated attention to neonates and hematology patients, two groups underrepresented in earlier validation work. It also complements the research group’s own prior multicenter study of the Phoenix score, pSOFA, and SIRS criteria in non-intensive care settings, extending the comparison to the sickest children in intensive care.

For the field, the message is one of calibrated confidence. The Phoenix sepsis score can now claim superior discrimination for in-hospital mortality in the general population of critically ill children with suspected infection, supporting its adoption in pediatric intensive care contexts. But the study simultaneously cautions against universal enthusiasm: in neonates, the older pSOFA remains an equally defensible choice, and in children with hematologic disease, both scores fall short of what clinicians would want from a decision-support tool, underscoring the need for population-specific approaches, whether through recalibration, integration of novel biomarkers, or entirely new models tailored to immunocompromised hosts. As sepsis mortality prediction moves from one-size-fits-all scores toward precision risk stratification, this study provides both a benchmark and a reminder that the sickest children often defy the averages on which general scores are built.

Subject of Research: External validation comparing the Phoenix sepsis score and pSOFA for predicting in-hospital mortality in critically ill children with suspected infection.

Article Title: Discrimination performance of the Phoenix sepsis score versus pediatric sequential organ failure assessment for predicting in-hospital mortality in critically ill children: a multicenter external validation study

Article References: Ru, X.-W., Liu, R.-Y., Yang, Z.-H., Huang, C.-Z., Zhang, C., Li, T.-W., Yang, J.-M., Liu, Y., Cui, X.-Y., Chen, H.-B., Zhou, X., Tang, F., Gong, X.-Y., Ling, P., Lv, X., Zeng, Q., Ma, Z., Yuan, B.-Y., Zhu, H., … Ye, Q. (2026). Discrimination performance of the Phoenix sepsis score versus pediatric sequential organ failure assessment for predicting in-hospital mortality in critically ill children: a multicenter external validation study. World Journal of Pediatrics. https://doi.org/10.1007/s12519-026-01073-y

Image Credits: AI Generated

DOI: 10.1007/s12519-026-01073-y

Keywords: Phoenix sepsis score, pSOFA, pediatric sepsis, in-hospital mortality, critical care, neonatal sepsis, hematology, AUROC, external validation, organ dysfunction, predictive scoring, intensive care unit