The tool, called SPACER, was developed by a large multi-institutional team led by researchers at UT Southwestern Medical Center, Baylor College of Medicine and UT MD Anderson Cancer Center. Its central innovation is a commitment to interpretability at a level that most machine learning systems abandon. Rather than burying its reasoning inside millions of opaque parameters, SPACER is built as a multi-instance learning framework that outputs gene-level recruitment rules directly from spatial transcriptomics data. In practical terms, a researcher can point the model at a tissue slide and receive back a ranked list of genes that either attract or repel a given cell type, with the logic connecting each gene to each cellular encounter laid bare for inspection and experimental testing.

The technical challenge the team confronted is worth appreciating. Cell recruitment in tissue is governed by an interplay of chemoattractants, adhesion molecules, extracellular matrix barriers, antigen presentation and countless other cues, all acting across microscopic distances. Standard computational approaches to spatial data tend to describe correlations between cell types and their surroundings, or to infer ligand–receptor communication from bulk expression patterns, but they rarely recover the causal grammar of recruitment. SPACER reframes the problem as one of learning from many small spatial instances—local patches of tissue in which a target cell either engaged with or was excluded from a particular neighborhood—and extracting the gene programs that consistently distinguish successful recruitment from exclusion. The result is a model whose internal logic maps onto biological mechanisms that experimenters can actually interrogate.

To establish credibility, the researchers benchmarked SPACER against existing methods across an unusually broad foundation of evidence: 37 spatially resolved transcriptomics datasets spanning multiple human tumor types as well as mouse heart tissue. Across a range of tasks designed to test whether a method can recover biologically meaningful recruitment signals, SPACER delivered improvements of fourfold to twenty-one-fold over prior approaches. That margin is not a marginal tuning gain; it reflects a qualitative difference in what the model can extract. The team also validated the framework against complementary data types, including single-cell RNA sequencing paired with T cell receptor sequencing, immunopeptidomics datasets cataloguing the peptide fragments displayed on cell surfaces, and newly generated Visium HD maps of mouse heart tissue.

The first major biological finding concerns mucins, the large, heavily glycosylated proteins best known for forming the protective mucus layer of airways and gut. SPACER’s analysis indicated that mucins expressed by tumor cells are associated with physical barriers that prevent T cells from engaging with the cancer cells around them. This is a mechanistically plausible picture: mucins can dramatically alter the biophysical properties of a cell’s surface and its surrounding matrix, and prior work has implicated them in shielding tumor cells from immune surveillance. What SPACER adds is a systematic, data-driven link between tumor mucin expression and the spatial exclusion of T cells across cancer types—a link that had been suspected but never decoded at this scale from spatial data.

The clinical implications of that finding are immediate. When the team examined cohorts of patients treated with immune checkpoint inhibitors—the blockbuster immunotherapy drugs that unleash T cells against tumors—mucin-associated recruitment barriers predicted resistance to treatment. The same signal also correlated with patient prognosis in pan-cancer analyses of The Cancer Genome Atlas, including in mesothelioma and kidney renal clear cell carcinoma. In other words, a computational readout of why T cells fail to reach tumor cells, derived purely from spatial gene expression maps, carries information that could in principle stratify patients before they embark on a course of therapy. That is precisely the kind of translational bridge that spatial biology has promised but rarely delivered.

The second headline discovery concerns what actually draws T cells into productive contact with tumor cells. Across cancer types, SPACER identified local antigen presentation as the dominant correlate of T cell engagement. This finding resonates with a core principle of immunology—T cells recognize their targets through peptide fragments displayed on major histocompatibility molecules—but seeing it emerge as the strongest spatial signal, outcompeting chemokine gradients and other candidate attractants, gives the principle new quantitative weight. The team reinforced the point with immunopeptidomics data, confirming that the peptides presented by tumor cells align with the T cell specificities inferred from receptor sequencing. Recruitment, in this view, is not a generic homing process but a antigen-informed negotiation between the immune cell and its immediate environment.

The third application took the team out of oncology and into a dangerous side effect of cancer therapy. Immune checkpoint inhibitors occasionally trigger myocarditis, a potentially fatal inflammation of the heart muscle, and which immune cells drive the damage has been debated. Applying SPACER to spatial maps of mouse heart tissue from a model of checkpoint inhibitor–induced myocarditis, together with single-cell RNA and T cell receptor sequencing, the researchers found that CD4-positive T cells—helper T cells more often associated with coordination than killing—are the dominant functional effectors in the severe form of the disease. Heart-infiltrating T cell clonotypes, expanded and largely distinct from those circulating in the blood, pointed to a local antigen-driven response within the myocardium, sharpening a picture previously suggested by work on α-myosin–reactive T cells.

Beyond these specific discoveries, the study sketches a broader vision for how interpretable machine learning could reshape tissue biology. The authors and their collaborators have made SPACER freely available through GitHub, and the underlying datasets—including human Visium and Visium HD data, gastric cancer CosMx maps, Xenium data paired with T cell receptor sequencing, and the mouse heart resources—are deposited in public repositories such as NIH GEO and Zenodo. That openness matters. A discovery engine for recruitment biology only earns trust if its rules can be independently tested, and the interpretability that defines SPACER is designed exactly for that cycle of computational prediction and experimental verification. The framework also proved computationally efficient, an unglamorous but essential quality for a tool intended to be applied across the rapidly growing volume of spatial datasets now being generated worldwide.

What makes the work resonate beyond its immediate findings is the way it reframes the relationship between artificial intelligence and biology. Foundation models for spatial omics are proliferating, and they are impressive at prediction, but their black-box character limits their scientific usefulness when the goal is mechanism. SPACER demonstrates the opposite philosophy: build interpretability in from the ground up, and the model becomes a hypothesis generator that hands experimenters concrete, testable molecular rules. The mucin barrier hypothesis, the centrality of local antigen presentation, and the CD4-driven pathology of immunotherapy myocarditis are all claims that can now be pursued in the laboratory—and, if they hold up, translated into biomarkers and therapeutic strategies. In decoding the spatial codes of cell recruitment, the study suggests that the most valuable AI in biomedicine may be the kind that shows its work.

Subject of Research: Interpretable machine learning for decoding cell recruitment rules from spatially resolved transcriptomics data

Article Title: Interpretable AI model reveals spatial codes of cell recruitment in tissues

Article References: Yao, J., Shim, J. S., Wong, K., Zhu, J., Liu, Y., Kang, S. W., McCord, K., Burt, B. M., Choi, J. M., Lee, C., Lubo Julio, I. C., Solis Soto, L. M., Hernandez, S., Colbert Maresso, K., Liu, Y., Dai, Y., Chu, T., Lu, W., Khan, K., … Wang, T. (2026). Interpretable AI model reveals spatial codes of cell recruitment in tissues. Nature Biomedical Engineering. https://doi.org/10.1038/s41551-026-01804-6

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01804-6

Keywords: spatial transcriptomics, interpretable AI, SPACER, T cell recruitment, tumor microenvironment, mucins, immune checkpoint inhibitors, antigen presentation, myocarditis, CD4 T cells, multi-instance learning, cancer immunotherapy

The clinical problem the researchers set out to solve is one that radiologists know all too well. Breast lesions classified as BI-RADS 4 or 5 — meaning suspicious enough to warrant biopsy — cover a wide spectrum of malignancy risk, and their ultrasound appearances overlap maddeningly. A benign lesion can look irregular in shape with indistinct margins and a dark, hypoechoic texture; so can a cancer. Posterior acoustic attenuation, a shadowing effect once considered a red flag, appears in both. This morphological ambiguity produces high inter-observer variability, particularly among less experienced practitioners. Ultrasound diagnosis is also highly operator-dependent, shaped by scanning skill, image acquisition standards, and subjective judgment. The stakes are high on both sides: a missed malignancy delays treatment, while an unnecessary core-needle biopsy or surgery exposes a patient to risk, anxiety, and cost for what turns out to be a harmless finding.

The challenge is compounded by breast tissue density. Asian women, on average, have denser breast tissue than other populations, which degrades the performance of mammography, the traditional screening mainstay. In China, clinical guidelines therefore pair mammography with ultrasound as the primary and auxiliary imaging methods for breast cancer detection. Ultrasound is convenient, radiation-free, and particularly effective in dense tissue — but its accuracy in distinguishing benign from malignant BI-RADS 4–5 lesions has remained stubbornly limited. Previous computational efforts largely focused on static images, requiring the sonographer or an algorithm to pick a single representative frame from a live scan. That choice discards a wealth of information: the way a lesion’s boundaries shift as the probe moves, how its internal echoes evolve, and the temporal context that an experienced human eye implicitly absorbs during a real-time examination.

TGMIL was designed specifically to exploit that discarded information. The architecture treats each ultrasound video as a bag of frames in a machine learning framework known as multi-instance learning, where the overall diagnosis is known but the diagnostic value of each individual frame is not. Each frame is first encoded by a VGG-13 convolutional backbone, generating a sequence of frame-level features. Crucially, the model also encodes the temporal order of frames through a linear layer, concatenating this positional information with the visual features so the network can learn from how a lesion appears, evolves, and disappears across the scan. An attention module then adaptively assigns higher weights to the most diagnostically informative frames and aggregates them into a single video-level representation. In effect, TGMIL learns which moments of the scan matter — suppressing redundant frames and emphasizing clinically relevant temporal segments — rather than treating every image equally.

The study enrolled 532 breast lesions categorized as BI-RADS 4–5 across four hospitals between November 2019 and January 2023. Of these, 334 lesions (61 benign and 273 malignant) formed the training and validation sets, while 198 lesions (50 benign and 148 malignant) were reserved as an independent test set. All videos were acquired with high-frequency linear probes, scanned continuously from the first appearance of each lesion to its complete disappearance, and converted to DICOM format. Preprocessing was rigorous: videos were converted to a lossless format, black borders and text overlays were stripped, a K-means algorithm extracted 15 key frames per video, and pixel values were standardized. Every diagnosis was confirmed by pathology from surgery or ultrasound-guided biopsy, and all examinations from the same patient were kept within a single dataset partition to prevent data leakage.

On the external test set, TGMIL achieved an area under the receiver operating characteristic curve of 0.824, decisively outperforming mainstream video classification architectures including 3D-CNN (AUC 0.665), CSN (0.613), and R(2+1)D-50 (0.678). It also beat static image-based classifiers: ResNet50, VGG16, the transformer-based ViT B/32 (which managed only 0.656), and even BU-Mamba, a recently proposed static ultrasound model. BU-Mamba achieved remarkable sensitivity of 0.988 but suffered from low precision of 0.752, indicating a tendency toward over-diagnosis. TGMIL struck a better balance, pairing high sensitivity of 0.966 with precision of 0.836 — meaning it reliably caught malignancies without flooding the clinic with false alarms. Ablation tests confirmed the value of the attention mechanism: simple mean pooling and max pooling baselines reached AUCs of only 0.752 and 0.770 respectively, showing that adaptively identifying informative frames, rather than averaging or maximizing across them, is what drives the performance gain.

The most clinically compelling results came from the reader study. Nine ultrasonographers — three juniors with under three years of experience, three attendings with four to nine years, and three seniors with a decade or more — independently reviewed the 198 test lesions, first without AI assistance and then, four weeks later and in a reshuffled case order, with TGMIL’s real-time benign-versus-malignant probability estimates. Unaided, the model itself achieved a patient-based AUC of 0.845, significantly outperforming both junior sonographers (0.682) and attendings (0.802). With TGMIL’s help, junior accuracy jumped from 66.7 to 89.2 percent and sensitivity surged from 67.4 to 94.2 percent, while attending accuracy rose from 80.8 to 90.0 percent. Across all readers, pooled accuracy improved from 77.5 to 95.0 percent, and the pooled AUC climbed from 0.871 to 0.964. Specificity for the whole group rose from 88.2 to 97.1 percent — a direct strike against unnecessary biopsies.

Interestingly, the most experienced readers gained the least. Senior ultrasonographers, who already achieved an AUC of 0.903 unaided, showed no statistically significant improvement on any metric with AI assistance, and their AUC dipped slightly from 0.898 to 0.885. The authors offer a candid explanation: because TGMIL’s standalone performance (0.824) falls below that of seasoned experts, the model may add little information beyond what veterans already extract from morphology, Doppler signals, and clinical intuition. Worse, when model predictions conflict with expert judgment, uncertain outputs can introduce cognitive interference or automation bias, occasionally nudging a correct decision astray. The study’s failure analysis reinforced this caution: some benign lesions displayed pseudo-malignant features such as spiculated margins that fooled the model into false positives, while some cancers with weak or atypical vascular signals escaped detection as false negatives.

The practical implications are nonetheless considerable. TGMIL was trained on a single NVIDIA RTX 3090 GPU — no supercomputer required — and once trained, inference on ultrasound video is computationally light enough for a local workstation, meaning the system could be integrated directly into the ultrasound workflow. The model also showed strong probability calibration (Brier score 0.071) and robust performance across all four external test centers, with AUCs ranging up to 0.886, suggesting it generalizes across different scanners, hospitals, and patient populations. For regions facing shortages of senior breast imaging specialists, an AI assistant that lifts junior sonographers to near-expert accuracy could meaningfully expand access to reliable diagnosis.

Limitations remain. The cohort was hospital-based and highly selected: only pathologically confirmed BI-RADS 4–5 lesions larger than 5 millimeters were included, excluding prostheses, poor-quality videos, and other BI-RADS categories. No asymptomatic screening population was tested, and screening settings — with lower disease prevalence and smaller incidental lesions — may behave very differently. The authors themselves urge prospective validation in screening cohorts before broader deployment. Still, the study marks a conceptual shift in medical AI: from judging isolated snapshots to understanding the full temporal narrative of a scan, the way a human expert does. If subsequent trials confirm these findings, the humble ultrasound video — long a passive record of a scan — may become an active diagnostic instrument, and the gap between novice and veteran sonographers may finally begin to close.

Subject of Research: Deep learning analysis of ultrasound videos for classifying benign and malignant BI-RADS 4–5 breast lesions

Article Title: AI-enhanced ultrasound video analysis for BI-RADS 4–5 breast lesions: a multicenter study of diagnostic performance

Article References: Li, W.-B., Du, Z.-C., Huang, W.-H., Qiu, R.-Y., Chen, Y.-C., Lu, L.-S., Ji, Z.-Y., Guo, L.-H., He, M.-Q., Wang, L.-S., & Zhang, G.-J. (2026). AI-enhanced ultrasound video analysis for BI-RADS 4–5 breast lesions: a multicenter study of diagnostic performance. Holistic Integrative Oncology, 5(1), Article 80. https://doi.org/10.1007/s44178-026-00302-8

Image Credits: AI Generated

DOI: 10.1007/s44178-026-00302-8

Keywords: artificial intelligence, deep learning, breast cancer, ultrasound, BI-RADS, computer-aided diagnosis, multi-instance learning, medical imaging, diagnostic accuracy, oncology, radiology, multicenter study

The clinical problem at the heart of the research is deceptively simple to state and maddeningly difficult to solve. Pneumonia-mimicking peripheral lung cancer, often abbreviated PLC, refers to lung tumors that arise in the outer regions of the lung and present on imaging as patchy opacities, ground-glass changes, or consolidation patterns that look almost indistinguishable from an infectious or inflammatory process. Radiologists confronting such lesions must weigh the possibility of malignancy against the far more common explanation of infection, and the stakes of that judgment are enormous. A false reassurance can allow an early-stage cancer to progress unchecked, while an unnecessary invasive workup exposes patients to procedural risk and anxiety. Conventional radiomics, which extracts hand-engineered quantitative features from medical images, has shown only limited diagnostic performance in this setting, leaving a genuine unmet need.

To close that gap, the research team assembled a dataset of 302 patients drawn from two medical centers, creating a multicenter cohort that lends the findings a degree of external validity often missing from single-institution machine learning studies. All patients had undergone venous-phase contrast-enhanced CT, the workhorse imaging protocol in which iodinated contrast agent highlights vascular and tissue enhancement patterns. Rather than feeding the network a single two-dimensional slice or the entire three-dimensional volume, the investigators adopted a middle path: they generated a 2.5D dataset by extracting seven consecutive slices centered on the largest axial cross-section of each lesion. This strategy captures contextual information above and below the most informative plane while keeping the data volume manageable, a pragmatic compromise between the simplicity of 2D analysis and the computational and data-hungry demands of full 3D modeling.

With the 2.5D data in hand, the team trained multiple deep learning architectures and compared their ability to separate cancer from mimicking inflammation. Among the candidates, ResNet101, a deep convolutional network whose residual connections allow very deep stacks of layers to be trained stably, emerged as the strongest performer, achieving an area under the receiver operating characteristic curve, or AUC, of 0.809 in the independent testing cohort, with a 95 percent confidence interval spanning 0.717 to 0.900. That figure, while respectable for a purely image-driven model, was only the beginning. The researchers then layered on more sophisticated machinery: multi-instance learning, a framework that treats each patient as a bag of image instances and learns from the collective evidence, and ensemble-based fusion strategies that combine the strengths of multiple models rather than betting on a single architecture.

The decisive leap in performance came when the deep learning features were fused with complementary sources of information. The team integrated predictive likelihood histogram features and bag-of-words features, techniques borrowed from natural language processing that summarize distributions of model outputs and image descriptors, with classical radiomics features and clinical variables. Statistical analysis identified three independent clinical predictors: patient sex, venous-phase CT attenuation values, and serum levels of carcinoembryonic antigen, a tumor marker frequently elevated in lung malignancies. When all of these strands were woven into a single combined model, the AUC rose to 0.874 in the training cohort and 0.855 in the independent testing cohort, the latter with a confidence interval of 0.774 to 0.936. In practical terms, the fused model correctly ranked cancerous lesions as more suspicious than their pneumonia lookalikes in roughly 85 to 87 percent of pairwise comparisons, a meaningful improvement over the image-only baseline.

Performance metrics alone, however, tell only part of the story, and the researchers were careful to interrogate their model with the full battery of validation tools now expected in clinical machine learning. Calibration curves demonstrated good agreement between predicted probabilities and observed outcomes, meaning that when the model expressed, say, a 70 percent likelihood of malignancy, roughly seven in ten such patients truly had cancer. Decision curve analysis, which quantifies the net clinical benefit of acting on a model’s predictions across a range of risk thresholds, showed that the combined model delivered higher net benefit than default strategies of treating all lesions as benign or all as malignant. These analyses matter because a model can achieve a dazzling AUC while being systematically overconfident or clinically useless at the thresholds where decisions are actually made, and the study’s results suggest the fused model avoids both traps.

Perhaps the most forward-looking aspect of the work is its commitment to interpretability, an attribute that has become a prerequisite for clinical acceptance of artificial intelligence. The team deployed two complementary explanation techniques. SHAP, or SHapley Additive exPlanations, draws on cooperative game theory to assign each input feature a quantified contribution to every individual prediction, revealing which clinical and imaging variables drove the model’s judgment in each case. Grad-CAM, or gradient-weighted class activation mapping, produces spatial heatmaps that highlight the image regions a convolutional network attended to when rendering its verdict. Together, these methods allow a radiologist to see not just what the model concluded but why, transforming an opaque algorithmic output into an auditable piece of diagnostic evidence that can be checked against human expertise.

The implications for clinical practice are considerable. A validated model of this kind could function as a silent second reader, flagging pneumonia-like lesions that warrant closer surveillance, PET imaging, or biopsy despite an ostensibly infectious appearance. Because the model relies on venous-phase contrast-enhanced CT, a scan already routinely obtained in the workup of indeterminate pulmonary lesions, it requires no new imaging protocol, no additional radiation exposure, and no extra cost beyond computation. The multicenter design, spanning the Second Affiliated Hospital of Nanchang University and the First Affiliated Hospital of Gannan Medical University, provides early reassurance that the model’s performance is not an artifact of one institution’s scanner fleet or patient mix, though prospective validation across broader and more diverse populations remains the essential next step before deployment.

The study also carries lessons for the wider field of medical artificial intelligence. Its architecture of success, combining a carefully engineered 2.5D input representation, multi-instance learning, ensemble fusion, and the deliberate integration of clinical laboratory data with imaging-derived features, illustrates that the best diagnostic models rarely come from raw images alone. The independent predictive value of carcinoembryonic antigen and venous-phase attenuation values underscores that biology and image intensity carry complementary signals, and that hybrid models can capture nuances invisible to either data stream in isolation. As screening programs expand and CT scans proliferate, tools that squeeze more diagnostic certainty from existing examinations will only grow in importance.

For now, the model remains a research instrument, its predictions confined to retrospective data and its authors appropriately cautious about immediate clinical use. Yet the trajectory is unmistakable. The work, supported by the Jiangxi Provincial Department of Science and Technology and conducted under the Declaration of Helsinki with ethics approval from Nanchang University, demonstrates that the pneumonia-mimicking lung cancer problem, long a source of diagnostic dread, is tractable to modern machine learning when the right data, architectures, and interpretability safeguards are brought to bear. If prospective studies confirm these results, the day may not be far off when every indeterminate, pneumonia-like spot on a chest CT is quietly accompanied by an algorithmic second opinion, one that has already learned to see through the disguise.

Subject of Research: Deep learning-based prediction of pneumonia-mimicking peripheral lung cancer on contrast-enhanced CT

Article Title: The predictive value of 2.5D deep learning model based on contrast-enhanced CT for pneumonia-mimicking peripheral lung cancer

Article References: Huang, H., Li, Q., Wang, J., Fu, X., Liu, Q., Liu, S., & Zuo, M. (2026). The predictive value of 2.5D deep learning model based on contrast-enhanced CT for pneumonia-mimicking peripheral lung cancer. BMC Cancer. https://doi.org/10.1186/s12885-026-17080-1

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17080-1

Keywords: peripheral lung cancer, pneumonia-mimicking lung cancer, deep learning, 2.5D model, contrast-enhanced CT, radiomics, multi-instance learning, ensemble learning, ResNet101, SHAP, Grad-CAM, machine learning

The research, led by Prudhvi Thirumalaraju, Manoj Kumar Kanakasabapathy, and Hadi Shafiee of Brigham and Women’s Hospital and Harvard Medical School, together with collaborators at Weill Cornell Medicine, introduces a model called Cohort-AI. Rather than evaluating embryos one by one, Cohort-AI uses a machine learning framework known as multi-instance learning to assess an entire patient’s set of embryos simultaneously, predicting the cumulative live-birth outcome of the cohort and ranking each embryo by its relative contribution to that outcome. The approach mirrors the way embryologists reason: an embryo that looks average on its own might be the best option in a weak cohort, while a good-looking embryo might rank lower in a set full of excellent candidates.

The motivation for the work comes from an uncomfortable finding in recent literature. A previous evaluation of eight commercial AI algorithms for embryo ranking found that agreement between the algorithms and human embryologists was lower than expected, and in some cases the models disagreed with each other. Most strikingly, two of the commercial models produced rankings that were statistically indistinguishable from random chance. Even when the models were fine-tuned on center-specific data, the discrepancies persisted. The problem, the researchers argue, is a phenomenon known in machine learning as underspecification: when many different internal solutions can achieve similar overall accuracy, small variations in training—such as the random seed used to initialize the network’s weights—can produce large, unpredictable changes in individual decisions, even while headline performance metrics look stable.

To test whether a cohort-aware design could tame this instability, the team trained fifty replicate versions of both a conventional single-instance convolutional neural network and the new Cohort-AI model, differing only in their random initialization. All training used more than 10,000 Day-5 embryo images from 1,258 patients at Massachusetts General Hospital, captured with time-lapse incubators. A separate dataset of 648 embryos from 53 patients at Weill Cornell was held out entirely as an independent external test set, with no tuning or adaptation of any kind. The researchers then measured how consistently the fifty replicate models of each type ranked the same embryos for the same patients.

The results were stark. On the Massachusetts General Hospital test set of 92 patient cohorts, the replicate Cohort-AI models achieved an average Kendall’s W—a statistical measure of rank-order agreement ranging from zero, meaning no agreement, to one, meaning perfect agreement—of approximately 0.94. The single-instance models managed only about 0.36. On the external Cornell data, the pattern held: roughly 0.91 for Cohort-AI versus 0.34 for the conventional approach, with both differences highly statistically significant. In practical terms, the conventional models, trained identically except for a random starting point, frequently disagreed with themselves about which embryo was best, while the cohort-aware models produced nearly identical rankings across all fifty replicates.

Consistency translated directly into fewer dangerous mistakes. The team defined a critical error as a case in which a model ranked a degenerate, grade 1 embryo as the top choice even though a viable blastocyst of grade 3 or better was present in the same cohort. On the internal test set, single-instance models made such errors at an average rate of about 12.4 percent, with individual models ranging from roughly 4 to 22 percent. Cohort-AI cut that rate to about 2.3 percent. On the external Cornell data the gap widened further: about 17.3 percent for the conventional models versus 1.24 percent for Cohort-AI. Notably, the variance in error rates across replicate models also collapsed for the cohort-aware approach, and formal testing showed that this variance did not significantly increase when the model was applied to the unfamiliar Cornell data—evidence that the stability survived a genuine distribution shift between two different clinics and microscope models.

The study also probed what the model had actually learned. Using attention mechanisms—the component of the network that assigns weights to different inputs when forming a decision—the researchers found that Cohort-AI consistently concentrated its attention on embryos of higher morphological grade, as judged by the modified Gardner grading system used in clinics, and assigned the lowest attention to empty wells in the culture dish. Feature-space visualizations showed that replicate Cohort-AI models clustered tightly together in how they represented embryos from the same patient, whereas the conventional models diverged widely. The authors are careful to note that attention weights are treated as a descriptive weighting mechanism, not as causal proof of embryo importance, but the alignment with embryologist grading suggests the model’s comparative reasoning resembles human judgment rather than exploiting spurious shortcuts.

Perhaps most importantly for clinical credibility, the stability did not come at the cost of performance. In retrospective comparisons against historical clinical outcomes, Cohort-AI’s top-ranked embryos matched the embryos clinicians actually transferred 68.7 percent of the time at Massachusetts General Hospital, compared with 44.7 percent for the single-instance models, and the live-birth rate associated with those top selections was 44.8 percent versus 41.0 percent—both above the center’s baseline of 35.1 percent for the evaluated patients. At Cornell, Cohort-AI again showed higher transfer and live-birth rates with at least twice the consistency. When the analysis was restricted to cycles known to contain at least one successful live birth, the median Cohort-AI model produced 29 live births from about 35 transfers on the internal data, while the median conventional model produced only 16 from about 23 transfers. The lowest-performing Cohort-AI replicates performed comparably to the highest-performing conventional models, meaning clinicians would no longer need to gamble on which replicate they happened to deploy.

A single-patient case study crystallized why these differences matter. For one patient with three high-quality blastocysts, one moderate blastocyst, and two degenerate embryos, consensus among five embryologists was unambiguous. Yet 33 of the 50 conventional models ranked a degenerate embryo above a high-quality one, 31 ranked a high-quality embryo last, and seven—including the three models with the highest validation accuracy—placed a degenerate embryo first. Pairwise rank correlations among the conventional models averaged just 0.27, with some pairs producing exactly opposite orderings. All fifty Cohort-AI models, by contrast, ranked the high-quality embryos at the top and the degenerate ones at the bottom, with an average pairwise correlation of 0.99. The two embryos Cohort-AI consistently ranked highest were among those that had resulted in live births. The lesson, the authors argue, is that strong validation accuracy is an insufficient proxy for clinical reliability in ranking tasks.

The implications extend beyond embryology. The authors contend that embryo selection is inherently a comparative problem, not a binary classification, and that validation frameworks designed for diagnostic AI—where each sample has a clear ground truth—are poorly suited to it. They call for standardized benchmarks that measure not only accuracy but also ranking reproducibility, cross-site robustness, and behavior across software updates, noting that AI adoption in IVF has surged from roughly a quarter of clinics in 2022 to more than half in 2025 even as many clinicians report low confidence in interpreting these systems. The team is candid about limitations: the study is retrospective, the associations with live birth are correlational rather than causal, both centers used the same family of time-lapse incubators, and no blinded comparison against embryologist rankings was performed. Prospective, multi-center trials remain the necessary next step. But the central message is clear and potentially field-changing: for AI to become dependable infrastructure in IVF, rankings must remain predictable as data and software evolve—and building the comparison into the model itself, rather than leaving it out, appears to be a powerful way to get there.

Subject of Research: A multi-instance learning AI model for stable and reliable AI-assisted embryo selection in IVF

Article Title: Cohort-AI: A multi-instance learning approach for improved stability and reliability in AI-assisted embryo selection

Article References: Thirumalaraju, P., Kanakasabapathy, M. K., Kandula, H., Kandula, T., Katkuri, A. V. R., Cipriano, C., Malmsten, J. E., Zaninovic, N., Bormann, C. L., & Shafiee, H. (2026). Cohort-AI: A multi-instance learning approach for improved stability and reliability in AI-assisted embryo selection. Heliyon, 12(15), Article e45457. https://doi.org/10.1016/j.heliyon.2026.e45457

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45457

Keywords: in vitro fertilization, embryo selection, artificial intelligence, multi-instance learning, deep learning, reproductive medicine, Kendall's W, model stability, underspecification, attention mechanisms, clinical decision support, time-lapse imaging