Researchers at New York University, alongside collaborators across seven countries, have developed and validated an artificial intelligence diagnostic tool capable of predicting breast cancer recurrence in hours rather than weeks. Published in Nature Communications in 2026, the study outlines a multimodal AI model that analyzes standard pathology slides alongside routine clinical data to estimate patient relapse risk. Led by NYU visiting scholar and Ataraxis AI co-founder Krzysztof J. Geras, with contributions from NYU computer science professor Yann LeCun, the system leverages self-supervised learning architecture DINOv2 to extract complex tissue patterns without relying on manually labeled datasets. Current standard practice relies on expensive genomic testing that requires weeks to generate results and consumes tumor tissue extracted during surgery, limiting options for subsequent analysis. The AI alternative bypasses these bottlenecks by reprocessing existing microscopic slides already examined by pathologists. In validation trials encompassing more than 3,500 patients across fifteen distinct cohorts, the model demonstrated high predictive accuracy, reliably stratifying individuals by recurrence risk. Notably, the AI successfully evaluated recurrence probabilities for hormone-receptor-positive, triple-negative, and HER2-positive breast cancer subtypes, addressing a significant gap in predictive tools for aggressive forms that currently lack reliable genomic assays. The researchers utilized standard statistical metrics, including the C-index and hazard ratios, to confirm the model discriminates effectively between high- and low-risk populations. According to Geras, the system matches or exceeds the performance of widely adopted genomic tests while delivering results in hours at a substantially lower cost. Crucially, by utilizing archived slides, the method preserves surgical tissue for future research or therapeutic testing. The authors emphasize that while the findings represent substantial progress in computational oncology, prospective randomized clinical trials are necessary to fully integrate the AI tool into standard treatment pathways. The underlying self-supervised framework, which learns rich biological representations prior to task-specific fine-tuning, holds potential for broader application across other disease domains, signaling a shift toward scalable, data-efficient AI in precision medicine.