Comprehensive evaluation of cross cancer generalization in histopathology segmentation models across 21 tumor types

Abstract

Background

Artificial intelligence has significantly advanced computational pathology by enabling high-resolution, clinical-grade tumor segmentation models with state-of-the-art diagnostic accuracy. Creating such models is resource-intensive, requiring substantial time and domain expertise. Additionally, deep learning models are typically restricted to single tumor types, making it challenging to develop separate models for each tumor type. Cross-cancer generalization of segmentation models could address this bottleneck and pave the way for pan-cancer segmentation models.

Methods

We evaluated the cross-tumor generalization capability of five tissue segmentation models (breast, colon, lung, kidney, prostate) using 21 cancer types from The Cancer Genome Atlas, totaling over 7,700 whole slide images. Representative large tumor and benign regions were manually selected, and segmentation accuracy was evaluated using a semiquantitative scale (0-10).

Results

Here we show that the lung model demonstrates excellent cross-cancer performance (overall mean score 7.9 ± 2.1), effectively segmenting tumor regions in many non-lung cancers with segmentation accuracy similar to its native domain in 11 of 19 other epithelial tumors and melanoma, achieving particularly strong results in ovarian cancer (9.2 ± 0.9). The breast and colon models also show strong cross-domain performance, while the kidney and prostate models exhibit more limited generalization. Overall, high-precision segmentation is achievable in most cancer types using existing models.

Conclusions

Existing segmentation models generalize across multiple cancer types, reducing the need to develop new, entity-specific models from scratch. This cross-domain generalization enables fast-track model development and supports future creation of robust pan-cancer segmentation models. Leveraging these capabilities could accelerate clinical integration of pathology artificial intelligence tools and enable reproducible biomarker discovery.

Plain language summary

Computer tools can help doctors who examine tissue samples under the microscope by automatically marking areas of cancer. However, building these tools for each type of cancer takes a lot of time and specialist knowledge. We asked whether a tool built for one type of cancer could also work for other types. We tested five such tools, each originally built for a specific cancer (breast, bowel, lung, kidney, or prostate), on tissue images from 21 different cancer types, covering more than 7,700 images in total. We found that several tools, especially the one built for lung cancer, could accurately identify cancer regions in many other cancer types as well. This means that existing tools can be reused for new cancers without having to build them from scratch, cutting the time needed from months to days and helping bring these computer-assisted tools to hospital laboratories sooner.