Dangott, Bryan

Foundation models are reshaping computational pathology by enabling scalable task-agnostic representations of histopathological whole-slide images (WSIs). Unlike earlier task-specific deep learning systems, pathology foundation models (PFMs) leverage massive whole-slide image repositories and self-supervised Vision Transformer architectures to achieve broad generalization and few-shot adaptability. Their evolution reflects a shift from weakly supervised approaches such as Clustering-Constrained Attention Multiple Instance Learning (CLAM) and hierarchical architectures such as Hierarchical Image Pyramid Transformer (HIPT) to large-scale efforts including foundation models, UNI, Virchow, Phikon, CONtrastive learning from Captions for Histopathology (CONCH), GigaPath, H-Optimus, Transformer-Based Pathology Image and Text Alignment Network (TITAN), and the Mayo Clinic Atlas. These models demonstrate impressive performance across diagnostic and prognostic benchmarks while also opening pathways for multimodal integration with genomics and clinical data. Yet significant barriers remain including inconsistent generalization across institutions, interpretability lagging behind clinical needs, and slow integration into routine laboratory workflows. Certain domains of anatomic pathology such as cytopathology, transplant pathology, frozen sections, and rare tumor subtypes remain particularly resistant to current models. Here, we review the development of PFMs, critically evaluate their strengths and limitations, and outline priorities for their safe and effective clinical translation. We argue that the next phase of PFM development will depend on rigorous benchmarking, pathologist-in-the-loop deployment, and multimodal fusion ensuring these models evolve from research tools into clinically robust systems.

Abstract Background: Recent advances in computer vision have given rise to image foundation models (FMs), large-scale neural networks pre-trained on millions of unlabeled histopathology patches, which can learn generalizable representations of tissue morphology. In many cancers, these FMs have predicted genetic mutations and prognoses directly from histology, suggesting that morphology encodes “hidden” molecular information. In lymphoma, this field is nascent but growing.General-purpose histopathology FMs have limited representation of lymphoid morphology. We developed and evaluated LymphoVision, a lymphoma-specialized image foundation model pretrained with self-supervised learning, to extract clinically meaningful features from routine hematoxylin and eosin (H&E) histology. Methods: LymphoVision was pre-trained using knowledge distillation and consistency (DINOv2) on over 37 million multi-resolution patches (5x, 10x, 20x, 40x) sampled using a density-aware strategy from 31,211 H&E whole slide images (WSIs) from 9,155 archival lymphoma and reactive lymphoproliferative cases. To ensure diverse, representative samples, our patch sampling method uses an efficient graph-based clustering algorithm and prioritizes morphologically informative regions by clustering patch embeddings and performing density sampling within each cluster. The pretrained encoder was evaluated using a clustering-constrained attention multiple instance learning (MIL) framework on three diagnostic tasks: (1) cell-of-origin (COO) classification in diffuse large B-cell lymphoma (DLBCL), distinguishing germinal center B-cell (GCB) from non-GCB subtypes based on the Hans immunohistochemistry algorithm; (2) follicular lymphoma (FL) grading, differentiating low-grade (grades 1–2) from high-grade (grade 3); and (3) multiclass lymphoma subtype classification across six categories: DLBCL, FL, marginal zone lymphoma (MZL), mantle cell lymphoma (MCL), classical Hodgkin lymphoma (CHL), and benign/reactive lymphoid conditions. WSI-level labels were derived from pathology reports, and none of the evaluation cases were included in pretraining. Models used a ViT-Giant LymphoVision backbone to generate 20x patch embeddings, followed by MIL using 50% training data and 2-fold cross-validation (25% validation, 25% test). Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, recall (sensitivity) and F1 score. Results: For DLBCL COO classification (total n=358 cases), LymphoVision achieved a mean test AUC of 0.93 (range 0.92–0.94), accuracy of 84.5% and F1 of 0.85 across two independent test folds (n=91 per fold, 50% GCB). These results outperformed general-purpose FMs on the same task and cases, including Virchow-v2 (AUC 0.85) and UNI-v2 (AUC 0.88), despite those models being pretrained on 10–100 times more images. For FL grading (total n=660 cases), the test set in each fold included 166 cases (18% grade 3), yielding a mean test AUC of 0.88 (range 0.87–0.90), accuracy of 89.2% and F1 of 0.80. In multiclass lymphoma subtype classification (total n=1,176 cases), the test set included 296 cases per fold. LymphoVision achieved a mean test AUC of 0.98 (range 0.97-0.99), mean test accuracy of 87.1% and mean F1 of 0.84. Class-wise recall was 92.0% for DLBCL (n=112), 88.2% for FL (n=72), 77.2% for MZL (n=29), 70.8% for MCL (n=35), 87.5% for CHL (n=24), and 95.8% for benign/reactive lymphoid conditions (n=24). Conclusion: LymphoVision is a lymphoma-specialized FM trained using self-supervision that achieves high diagnostic accuracy across multiple clinically meaningful tasks using weak supervision and standard H&E images, suggesting the encoder has learned fine-grained lymphoma morphology. Based on historical benchmarks from the 1997 International Lymphoma Study Group classification project, expert hematopathologists achieved diagnostic agreement rates of 55% to 84% using H&E morphology alone for the non-Hodgkin lymphoma (NHL) entities included in our multiclass task. LymphoVision approached or exceeded these benchmarks. These findings underscore the value of our approach using disease-specialized pretraining and intelligent patch sampling for data-efficient histopathology learning. LymphoVision has the potential to augment diagnostics, enable scalable AI-powered classification in real-world workflows, and serve as a foundation for prognostic and theragnostic applications. External validation and multimodal integration are ongoing.

Abstract Background Thyroid function testing is a common laboratory service that can be optimized by implementing evidence-based algorithms. The thyroid function cascade testing approach at Mayo Clinic Labs is utilized to effectively evaluate and monitor functional thyroid status. The cascade begins with thyroid-stimulating hormone (TSH) as a screening marker. When TSH levels are normal, no additional testing is necessary. When the TSH result is abnormal, appropriate follow-up free T4, total T3, and/or TPO testing will automatically be performed. Providers at our institution can order thyroid function testing with this cascade testing approach or as standalone individual orderable tests. Initial data analysis showed patterns of TSH ordered concurrently with free T4, total T3, and/or anti-TPO. Further analysis revealed that a high number of concurrently ordered tests would not have been necessary had the algorithmic approach been followed. This opportunity for improvement led to the thyroid function testing initiative to direct providers to order the cascade test rather than individual thyroid tests, thereby reducing concurrent and unnecessary testing to improve patient care. Methods The initiative involved creating a new Epic order panel directing providers to order the cascade rather than the individual orderable tests. This intervention is a soft stop that allows providers to proceed by ordering individual tests if desired. We evaluated the impact of the initiative on test utilization, cost savings, and quality indicators using pre- and post-intervention data from all Mayo Clinic sites over a 9-month period. We compared the monthly volumes of concurrent and unnecessary orders of Free T4, Total T3, and TPO before and after the intervention. Concurrent orders were defined as tests collected from the same patient on the same date and time as TSH tests. Unnecessary orders were defined as tests that were outside the reflex range according to the thyroid function ordering algorithm. Cost savings were calculated based on the 2024 CMS reimbursement rates for the associated test CPT® codes. We also monitored and addressed feedback and requests from providers and laboratory staff throughout the initiative. Results The intervention resulted in a 20% average monthly reduction in concurrent orders and an 18% average monthly reduction in unnecessary orders of free T4, total T3, and anti-TPO tests. Estimated savings based on 2024 CMS reimbursement rates are $5,000 monthly and nearly $60,000 annually. Thyroid testing quality was also improved by aligning with clinical guidelines, reducing over-testing, and minimizing the risk of misdiagnosis. Conclusion The thyroid function testing initiative was a successful laboratory stewardship project that reduced redundant testing, improved patient care, and reduced costs. The initiative demonstrated the feasibility and effectiveness of using evidence-based reflex algorithms to optimize laboratory testing. The lessons learned include assessing all tests related to the disease condition to mitigate potential unintended increases in test usage and accounting for patient groups that are not covered by the intended algorithm*s target population. Future directions include expanding the cascade approach to other areas of laboratory medicine, evaluating the long-term outcomes of the initiative, and engaging with providers to ensure the adoption of the initiative.

Artificial Intelligence (AI) has the potential to revolutionize cytopathology by enhancing diagnostic accuracy, efficiency, and accessibility. However, the implementation of AI in this field presents significant challenges and opportunities. This review paper explores the current landscape of AI applications in cytopathology, highlighting the critical challenges, including data quality and availability, algorithm development, integration and standardization, and clinical validation. We discuss challenges such as the limitation of only one optical section and z-stack scanning, the complexities associated with acquiring high-quality labeled data, the intricacies of developing robust and generalizable AI models, and the difficulties in integrating AI tools into existing laboratory workflows. The review also identifies substantial opportunities that AI brings to cytopathology. These include the potential for improved diagnostic accuracy through enhanced detection capabilities and consistent, reproducible results, which can reduce observer variability. AI-driven automation of routine tasks can significantly increase efficiency, allowing cytopathologists to focus on more complex analyses. Furthermore, AI can serve as a valuable educational tool, augmenting the training of cytopathologists and facilitating global health initiatives by making high-quality diagnostics accessible in resource-limited settings. The review underscores the importance of addressing these challenges to harness the full potential of AI in cytopathology, ultimately improving patient care and outcomes.

Transplant pathology plays a critical role in ensuring that transplanted organs function properly and the immune systems of the recipients do not reject them. To improve outcomes for transplant recipients, accurate diagnosis and timely treatment are essential. Recent advances in artificial intelligence (AI)-empowered digital pathology could help monitor allograft rejection and weaning of immunosuppressive drugs. To explore the role of AI in transplant pathology, we conducted a systematic search of electronic databases from January 2010 to April 2023. The PRISMA checklist was used as a guide for screening article titles, abstracts, and full texts, and we selected articles that met our inclusion criteria. Through this search, we identified 68 articles from multiple databases. After careful screening, only 14 articles were included based on title and abstract. Our review focuses on the AI approaches applied to four transplant organs: heart, lungs, liver, and kidneys. Specifically, we found that several deep learning-based AI models have been developed to analyze digital pathology slides of biopsy specimens from transplant organs. The use of AI models could improve clinicians’ decision-making capabilities and reduce diagnostic variability. In conclusion, our review highlights the advancements and limitations of AI in transplant pathology. We believe that these AI technologies have the potential to significantly improve transplant outcomes and pave the way for future advancements in this field.

Abstract Aims Current non-invasive screening methods for cardiac allograft rejection have shown limited discrimination and are yet to be broadly integrated into heart transplant care. Given electrocardiogram (ECG) changes have been reported with severe cardiac allograft rejection, this study aimed to develop a deep-learning model, a form of artificial intelligence, to detect allograft rejection using the 12-lead ECG (AI-ECG). Methods and results Heart transplant recipients were identified across three Mayo Clinic sites between 1998 and 2021. Twelve-lead digital ECG data and endomyocardial biopsy results were extracted from medical records. Allograft rejection was defined as moderate or severe acute cellular rejection (ACR) based on International Society for Heart and Lung Transplantation guidelines. The extracted data (7590 unique ECG-biopsy pairs, belonging to 1427 patients) was partitioned into training (80%), validation (10%), and test sets (10%) such that each patient was included in only one partition. Model performance metrics were based on the test set (n = 140 patients; 758 ECG-biopsy pairs). The AI-ECG detected ACR with an area under the receiver operating curve (AUC) of 0.84 [95% confidence interval (CI): 0.78–0.90] and 95% (19/20; 95% CI: 75–100%) sensitivity. A prospective proof-of-concept screening study (n = 56; 97 ECG-biopsy pairs) showed the AI-ECG detected ACR with AUC = 0.78 (95% CI: 0.61–0.96) and 100% (2/2; 95% CI: 16–100%) sensitivity. Conclusion An AI-ECG model is effective for detection of moderate-to-severe ACR in heart transplant recipients. Our findings could improve transplant care by providing a rapid, non-invasive, and potentially remote screening option for cardiac allograft function.

Abstract Background Current approaches utilizing non-invasive methods to screen for cardiac allograft rejection (gene expression profiling and cell free DNA) have yet to be broadly integrated into heart transplant management and have shown limited discrimination (AUCs of 0.68 and 0.77, respectively). Changes in the electrocardiogram (ECG) have been reported at the time of severe cardiac rejection, including low voltages and conduction abnormalities. It remains unknown if subtle ECG changes correlating with cardiac allograft rejection can be detected earlier using machine learning methods. Purpose We sought to develop an artificial intelligence (AI) model to detect cardiac allograft rejection based on the 12 lead ECG. Methods We identified all patients who underwent a heart transplant at 3 hospital sites within a single health system from Jan 1998 through Apr 2021 and extracted digital 12-lead ECG data as well as endomyocardial biopsy pathology results from the electronic medical record. We partitioned our data into a training (80%), validation (10%), and test set (10%) based on a group stratification sampling. Each patient was present in only one set and each set had a positivity rate of 2.6% with 6,074/758/758 ECGs belonging to 1,146/140/141 unique patients in each set respectively. Cardiac allograft rejection was defined as moderate or severe acute cellular rejection based on International Society for Heart and Lung Transplantation (ISHLT) guidelines. A convolutional neural network, using the 12-lead ECG data as input, was trained with hyperparameter optimization for regularization, learning rate adjustments, and class weights. Model performance metrics were based on the test data and estimated using the final model architecture. Results 1,587 heart transplant recipients who had at least one endomyocardial biopsy were evaluated for inclusion. We limited our sample to ECGs performed within 30 days of the biopsy date (7,590 ECGs, representing 1,425 unique patients). Our study population had a median age of 55.8 years and 28.7% were female. The median number of ECG-biopsy pairs per patient was 5. The majority of endomyocardial biopsy results were classified as none or mild rejection (97.1%), and 2.9% had moderate/severe rejection. The ECG-based AI model detected cardiac allograft rejection with an area under the receiver operative curve (AUC) of 0.84 in the test set. The sensitivity, specificity, positive and negative predictive values were 95%, 52.6%. 5.2% and 99.7% respectively. Conclusions An AI-ECG model appears to outperform novel non-invasive laboratory tests (gene expression profiling and cell free DNA) for detecting cardiac allograft rejection and does not require a blood draw or the additional complexities surrounding sample processing. This model relies on a readily available and relatively inexpensive test, the ECG. In addition, AI predictions can be made available within a few minutes following ECG acquisition. Funding Acknowledgement Type of funding sources: Private hospital(s). Main funding source(s): Mayo Clinic