Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generalisation of automatic tumour segmentation in histopathological whole-slide images across multiple cancer types

Ole-Johan Skrede, Manohar Pradhan, Maria Xepapadakis Isaksen, Tarjei Sveinsgjerd Hveem, Ljiljana Vlatkovic, Arild Nesbakken, Kristina Lindemann, Gunnar B Kristensen, Jenneke Kasius, Alain G Zeimet, Odd Terje Brustugun, Lill-Tove Rasmussen Busund, Elin H Richardsen, Erik Skaaheim Haug, Bjørn Brennhovd, Emma Rewcastle, Melinda Lillesand, Vebjørn Kvikstad, Emiel Janssen, David J Kerr, Knut Liestøl, Fritz Albregtsen, Andreas Kleppe

TL;DR

The study demonstrates that a single pan-cancer segmentation model can automatically delineate tumour regions in H&E WSIs across colorectal, endometrial, lung, and prostate cancers, with external validation in breast and bladder cohorts achieving mean $DSC$ values over 80% and often above 90% for endometrial cancer. Trained on over 20,000 WSIs from more than 4,000 patients, the universal model matches or exceeds the performance of cancer-type–specific models and remains robust to scanner differences and variable sample preparation, including TCGA datasets. Limitations include reduced performance for small, fragmented, or early-stage bladder tumours and potential biases from single-pathologist annotations and resections rather than biopsies. The findings support deploying pan-cancer segmentation as a reliable first step in digital pathology pipelines, enabling scalable analysis of tumour areas across diverse patient populations and technical workflows.

Abstract

Deep learning is expected to aid pathologists by automating tasks such as tumour segmentation. We aimed to develop one universal tumour segmentation model for histopathological images and examine its performance in different cancer types. The model was developed using over 20 000 whole-slide images from over 4 000 patients with colorectal, endometrial, lung, or prostate carcinoma. Performance was validated in pre-planned analyses on external cohorts with over 3 000 patients across six cancer types. Exploratory analyses included over 1 500 additional patients from The Cancer Genome Atlas. Average Dice coefficient was over 80% in all validation cohorts with en bloc resection specimens and in The Cancer Genome Atlas cohorts. No loss of performance was observed when comparing the universal model with models specialised on single cancer types. In conclusion, extensive and rigorous evaluations demonstrate that generic tumour segmentation by a single model is possible across cancer types, patient populations, sample preparations, and slide scanners.

Generalisation of automatic tumour segmentation in histopathological whole-slide images across multiple cancer types

TL;DR

The study demonstrates that a single pan-cancer segmentation model can automatically delineate tumour regions in H&E WSIs across colorectal, endometrial, lung, and prostate cancers, with external validation in breast and bladder cohorts achieving mean values over 80% and often above 90% for endometrial cancer. Trained on over 20,000 WSIs from more than 4,000 patients, the universal model matches or exceeds the performance of cancer-type–specific models and remains robust to scanner differences and variable sample preparation, including TCGA datasets. Limitations include reduced performance for small, fragmented, or early-stage bladder tumours and potential biases from single-pathologist annotations and resections rather than biopsies. The findings support deploying pan-cancer segmentation as a reliable first step in digital pathology pipelines, enabling scalable analysis of tumour areas across diverse patient populations and technical workflows.

Abstract

Deep learning is expected to aid pathologists by automating tasks such as tumour segmentation. We aimed to develop one universal tumour segmentation model for histopathological images and examine its performance in different cancer types. The model was developed using over 20 000 whole-slide images from over 4 000 patients with colorectal, endometrial, lung, or prostate carcinoma. Performance was validated in pre-planned analyses on external cohorts with over 3 000 patients across six cancer types. Exploratory analyses included over 1 500 additional patients from The Cancer Genome Atlas. Average Dice coefficient was over 80% in all validation cohorts with en bloc resection specimens and in The Cancer Genome Atlas cohorts. No loss of performance was observed when comparing the universal model with models specialised on single cancer types. In conclusion, extensive and rigorous evaluations demonstrate that generic tumour segmentation by a single model is possible across cancer types, patient populations, sample preparations, and slide scanners.

Paper Structure

This paper contains 39 sections, 3 equations, 8 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Materials
  4. Primary analysis of model performance
  5. Factors affecting segmentation performance
  6. Comparison with models specialised on single cancer types
  7. Robustness to variations in sample origin, preparation and imaging
  8. Intra- and inter pathologist variability
  9. Failure to segment fragmented samples from early-stage tumours in bladder
  10. Performance comparison with MedSAM
  11. Discussion
  12. Methods
  13. Materials
  14. Development cohorts
  15. Validation cohorts
  16. ...and 24 more sections

Figures (8)

  • Figure 1: Overview of input and corresponding result
  • Figure 2: Included patient and WSI count
  • Figure 3: Segmentation method pipeline
  • Figure 4: Example result in TCGA-FD-A6TE-01Z-00-DX1 from BLCA
  • Figure 5: Primary model results in cohorts from development, validation, and TCGA
  • ...and 3 more figures