Pathology Slides Reveal More Than a Tumor’s Appearance: AI Identifies Microregions Associated With Recurrence Risk in Triple-Negative Breast Cancer

Triple-negative breast cancer lacks common therapeutic targets such as the estrogen receptor, progesterone receptor, and HER2. Even when patients receive similar treatments, their risks of recurrence can still differ substantially. A new study started with H&E-stained slides already produced during routine pathology examinations, using artificial intelligence not only to predict risk but also to identify locations within tumors that may harbor molecular states associated with recurrence.

The research team analyzed slides from 156 patients, training a model to score local regions in the images and then aggregate the distribution of high-scoring regions into patient-level recurrence risk. In an independent test set, both the model’s AUC and C-index were 0.77. The former measures the model’s ability to distinguish between risk groups, while the latter reflects the agreement between predicted rankings and actual times to recurrence. These results indicate that the model has moderate discriminatory ability, but not enough to independently support clinical decision-making.

More biologically significant was the heatmap finding that high- and low-risk regions could coexist within the same tumor tissue compartment. Global proteomic analysis linked imaging-defined high risk to processes such as the cell cycle and genome maintenance, while low risk was more closely associated with immune activation. In other words, the AI may have captured not only how tumor cells are arranged, but also distinct proliferative and immune environments underlying the slide morphology.

To further examine this hypothesis, the team used coordinates from the AI heatmaps to isolate 46 regions from the tumors of two patients who had experienced recurrence and performed mass spectrometry-based spatial proteomic analysis. The two patients showed similar contrasts: high-risk regions were enriched for mitosis-related processes, whereas low-risk regions showed more immune and antigen-presentation activity. This step connected the imaging signal to local molecular mechanisms, but because the samples came from only two patients, it is more appropriately regarded as mechanistic exploration than broadly applicable clinical evidence.

The researchers also used these spatial differences to develop a composite 13-protein index. In an expanded cohort, the protein-based version showed only a trend in which higher scores were associated with worse recurrence-free survival; the corresponding transcript-based index, however, distinguished recurrence-free survival in an independent METABRIC triple-negative breast cancer cohort. In another resampling analysis, adding the protein index to the H&E imaging score increased the C-index from 0.679 to 0.739 and also improved time-dependent discrimination at three and five years.

For now, the most immediate value of this approach may not be generating a new automated interpretation score, but using AI heatmaps as an experimental navigation tool: first identifying suspicious regions on inexpensive, widely available slides, then concentrating costly molecular analyses on the locations with the highest information content. However, the paper has not yet undergone peer review, the spatial proteomics component was very small, and the prognostic association of the protein index remains preliminary. Before clinical use, the approach will require validation across more medical institutions, different scanning and staining workflows, and prospective patient data, as well as evidence that it provides consistent added value beyond existing pathology and staging information.