Reproduce “Generalisation of automatic tumour segmentation in histopathological whole-slide images across multiple cancer types” on OpenResearch

About this paper

Ole-Johan Skrede, Manohar Pradhan, Maria Xepapadakis Isaksen, Tarjei Sveinsgjerd Hveem, Ljiljana Vlatkovic, Arild Nesbakken, Kristina Lindemann, Gunnar B Kristensen and 15 more

A universal deep learning model for tumour segmentation in histopathological whole-slide images demonstrates robust generalization across multiple common and previously unseen cancer types, various patient populations, sample preparations, and scanner models. The model consistently achieved a mean Dice Similarity Coefficient exceeding 80% across external validation cohorts, performing comparably to specialized single-cancer models and within the range of human pathologist variability.

Key insights

• A single deep learning model can effectively learn generalizable characteristics of cancerous tissue morphology, enabling robust tumour segmentation across diverse cancer types, patient populations, and technical variations without performance compromise.
• Extensive and varied training data, coupled with pre-planned, rigorous external validation across different institutions and scanner technologies, is critical for developing clinically trustworthy and broadly applicable AI models in pathology.
• The performance of an automated tumour segmentation model can rival or exceed the consistency of human expert agreement, providing a reliable and scalable benchmark for diagnostic tasks.

Summary by alphaXiv.