Pathology

Existing segmentation models generalize across many cancer types, with variable performance.

Five pretrained tissue segmentation models were tested on 7,700 TCGA slides spanning 21 cancer types to assess cross-cancer generalization. The lung model achieved a mean semiquantitative score of 7.9 ± 2.1 and matched native-domain performance in 11 of 19 epithelial non-lung tumors. Breast and colon models also generalized well, while kidney and prostate models showed limited cross-domain accuracy. Clinical implication: laboratories can consider repurposing high-performing models to accelerate pan-cancer segmentation but must audit performance by tumor type first. 1

Deep learning on CD3 slides stratifies prognosis in stage III colon cancer across international cohorts.

A VGG19-based model analyzed CD3-stained slides from 1,737 stage III colon cancer patients in three cohorts to identify tumor core and invasive margin features. The model clustered patients into groups with significantly different disease-free survival across training and validation sets (IM: p < 0.001 in training; external p = 0.02). Combined classifiers improved prognostic accuracy and outperformed traditional clinical variables and manual CD3 counts. Clinical implication: automated CD3 image analysis can augment prognostic stratification but needs pathway integration and prospective validation before guiding treatment. 2

Digepath: a GI-focused foundation model with broad task performance and workflow integration.

Digepath was pretrained on 353 million patches from 210,043 H&E slides and fine-tuned on 471,443 expert-annotated regions for GI pathology. It achieved state-of-the-art results on 32 of 33 downstream GI tasks including diagnosis, molecular profiling, and prognosis. The model is integrated into an agent-based clinical reasoning framework to support end-to-end diagnostic workflows. Clinical implication: Digepath may improve GI diagnostic reproducibility, but local validation and workflow testing are required before deployment. 3