A histomorphological atlas of resected mesothelioma discovered by self-supervised learning from 3446 whole-slide images

Our plain-English summary of A histomorphological atlas of resected mesothelioma discovered by self-supervised learning from 3446 whole-slide images, by F. Seyedshahi, K. Rakovic, N. Poulain, et al (Nature Communications, 2025).

The study

• Authors: F. Seyedshahi, K. Rakovic, N. Poulain, A. Claudio Quiros, I.R. Powley, C. Richards, H. Uraiby, S. Klebe, D.A. Moore, A. Nakas, C.R. Wilson, M. Sereno, L. Officer-Jones, C. Ficken, A. Teodosio, F. Ballantyne, D. Murphy, K. Yuan, J. Le Quesne
• Journal: Nature Communications
• Published: 2025;16(1):8891
• Cited by: 2 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Mesothelioma is a highly lethal and poorly biologically understood disease which presents diagnostic challenges due to its morphological complexity. This study uses self-supervised AI (Artificial Intelligence) to map the histomorphological landscape of the disease. The resulting atlas consists of recurrent patterns identified from 3446 Hematoxylin and Eosin (H&E) stained images scanned from resected tumour slides. These patterns generate highly interpretable predictions, achieving state-of-the-art performance with 0.65 concordance index (c-index) for outcomes and 88% AUC in subtyping. Their clinical relevance is endorsed by comprehensive human pathological assessment. Furthermore, we characterise the molecular underpinnings of these diverse, meaningful, predictive patterns. Our approach both improves diagnosis and deepens our understanding of mesothelioma biology, highlighting the power of this self-learning method in clinical applications and scientific discovery.

© 2025. The Author(s).

In plain English

This study used a self-supervised artificial intelligence method to sort microscope slide images of surgically removed pleural mesothelioma into recurring tissue patterns. The main dataset was a Leicester, UK archive (LATTICe-M) of 512 patients. After exclusions, 3446 slide images from 485 patients were analysed. The software found 47 patterns. Three expert pathologists then described them, with fair to moderate agreement among themselves. Using these patterns, the model told epithelioid tumours apart from non-epithelioid (sarcomatoid or biphasic) tumours with an AUC of 88% in the Leicester data and 80% in an outside dataset (TCGA). For predicting survival, its concordance index was 0.65 in Leicester test data and 0.65 in TCGA. Patterns labelled "epithelioid nests in bland stroma" and "dense lymphocytes" were associated with better survival. "Disorderly spindle cells" and "transitional mesothelioma" were associated with poorer survival. In TCGA epithelioid cases, the model's risk groups differed in survival (p = 0.025), while groups based on standard human grading did not differ significantly (p = 0.282). The authors concluded the method achieved state-of-the-art accuracy. They noted that staging data were missing for 32.23% of Leicester cases and smoking history was incomplete for 43.55%.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Retrospective computational (self-supervised artificial intelligence) study of archived pathology slide images, tested on two outside datasets
• Who was studied: 512 patients with pleural mesothelioma who had surgical resection, from the Leicester LATTICe-M archive (372 epithelioid, 107 sarcomatoid, 33 biphasic; 436 male, 76 female; aged 36 to 85). After excluding slides with too little tissue, 485 patients and 3446 slide images remained. Outside datasets were TCGA (86 samples from 74 patients) and 235 tissue cores from St. George's Hospital, London.
• Main results:
  • 47 recurring tissue-pattern clusters were identified from 3,239,939 image tiles taken from 3446 slide images.
  • Agreement among the three pathologists describing the patterns (Fleiss' Kappa) ranged from 0.2 to 0.6, which the authors called fair to moderate.
  • Classifying epithelioid versus non-epithelioid (sarcomatoid/biphasic) subtype: 88% 5-fold cross-validated AUC in the LATTICe-M data.
  • Subtype classification AUC was 80% in the TCGA mesothelioma dataset.
  • Survival prediction: 5-fold cross-validated concordance index (c-index) of 0.67 in LATTICe-M training data and 0.65 in LATTICe-M test data.
  • Survival prediction c-index was 0.65 in the fully unseen TCGA cohort.
  • Adding clinical information (subtype, TNM stage, age) improved the c-index by 0.01.
  • Two patterns were associated with better survival. HPC 10, "epithelioid nests in bland stroma": log hazard ratio -0.089, p = 0.001, confidence interval -0.145 to -0.034. HPC 27, "dense lymphocytes": log hazard ratio -0.062, p = 0.008, confidence interval -0.109 to -0.016.
  • Two patterns were described as predictors of poor outcome. HPC 15, "disorderly spindle cells": log hazard ratio 0.052, p = 0.026, confidence interval 0.006 to 0.098. HPC 22, "transitional mesothelioma": log hazard ratio 0.042, p = 0.016, confidence interval 0.008 to 0.077.
  • In TCGA epithelioid cases, the model's high- and low-risk groups differed in survival (log-rank p = 0.025). Groups based on classical human grading did not differ significantly (p = 0.282).
  • In TCGA gene-expression data, spindle cell-enriched patterns (HPCs 22, 16, 15 and 6) showed especially high proliferation.
  • The lymphocyte-rich pattern HPC 27 correlated with T cell, B lineage cell and myeloid dendritic cell signatures.
  • On 235 St. George's Hospital tissue cores, the model's multiple-instance version (HPL-MIL) had higher AUC, average precision, sensitivity and specificity than the other methods compared. Its AUC was 0.93 plus or minus 0.12, versus 0.90 for MesoGraph.
• The authors' conclusion: The authors concluded that their method achieved state-of-the-art accuracy in survival prediction and subtype classification, and outperformed human grading for prognosis in epithelioid cases.
• Limitations the authors state:
  • Staging data were missing for 32.23% of Leicester cases (165 patients), probably reducing the power of stage-related analysis.
  • Smoking history was incomplete for 43.55% of cases (223 patients), limiting assessment of its impact.

Topics

The National Library of Medicine indexes this study under subjects our site covers: Mesothelioma.

Medical terms in this study, explained in our Lexicon: Mesothelioma, Concordance Index, Discovery.

Related studies

• Predicting cancer using supervised machine learning: Mesothelioma — A. Choudhury (2021)
• Accelerating Whole-Body Diffusion-weighted MRI with Deep Learning-based Denoising Image Filters — K. Zormpas-Petridis, N. Tunariu, et al. (2021)
• Cancer-Associated Fibroblasts Influence Survival in Pleural Mesothelioma: Digital Gene Expression Analysis and Supervised Machine Learning Model — S. Borchert, A. Mathilakathu, et al. (2023)
• Asbestos Detection with Fluorescence Microscopy Images and Deep Learning — C. Cai, T. Nishimura, et al. (2021)
• Transfer learning approach in pre-treatment CT images to predict therapeutic response in advanced malignant pleural mesothelioma — A. Fanizzi, A. Catino, et al. (2024)
• Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images — S.C. Baek, K.H. Lee, et al. (2023)

Browse more research in Nature Communications, or return to medical research.