Deep learning-based classification of mesothelioma improves prediction of patient outcome

Our plain-English summary of Deep learning-based classification of mesothelioma improves prediction of patient outcome, by P. Courtiol, C. Maussion, M. Moarii, et al (Nature Medicine, 2019).

The study

• Authors: P. Courtiol, C. Maussion, M. Moarii, E. Pronier, S. Pilcer, M. Sefta, P. Manceron, S. Toldo, M. Zaslavskiy, N. Le Stang, N. Girard, O. Elemento, A.G. Nicholson, J.Y. Blay, F. Galateau-Sallé, G. Wainrib, T. Clozel
• Journal: Nature Medicine
• Published: 2019;25(10):1519-1525
• Cited by: 335 articles, as counted by NIH iCite
• Read it: PubMed · Publisher

Abstract

Malignant mesothelioma (MM) is an aggressive cancer primarily diagnosed on the basis of histological criteria1. The 2015 World Health Organization classification subdivides mesothelioma tumors into three histological types: epithelioid, biphasic and sarcomatoid MM. MM is a highly complex and heterogeneous disease, rendering its diagnosis and histological typing difficult and leading to suboptimal patient care and decisions regarding treatment modalities2. Here we have developed a new approach-based on deep convolutional neural networks-called MesoNet to accurately predict the overall survival of mesothelioma patients from whole-slide digitized images, without any pathologist-provided locally annotated regions. We validated MesoNet on both an internal validation cohort from the French MESOBANK and an independent cohort from The Cancer Genome Atlas (TCGA). We also demonstrated that the model was more accurate in predicting patient survival than using current pathology practices. Furthermore, unlike classical black-box deep learning methods, MesoNet identified regions contributing to patient outcome prediction. Strikingly, we found that these regions are mainly located in the stroma and are histological features associated with inflammation, cellular diversity and vacuolization. These findings suggest that deep learning models can identify new features predictive of patient survival and potentially lead to new biomarker discoveries.

In plain English

The authors built and tested MesoNet, a deep learning computer program that predicts survival of mesothelioma patients from digitized tissue slides alone. They trained it on slides from 2,300 pleural mesothelioma patients in the French MESOPATH/MESOBANK database, tested it on 681 others, and checked it on 56 patients from The Cancer Genome Atlas (TCGA). Asbestos exposure was not studied. On a ranking measure called the c index, MesoNet scored 0.643 in the test set, compared with 0.598 for the pathologist-assigned tumor type. In the TCGA set it scored 0.656, compared with 0.590. Of the 60 test patients it predicted would fare worst, 34% had sarcomatoid tumors and 40% had epithelioid tumors. Their outcomes were comparable to those of the 60 sarcomatoid patients. Two pathologists reviewed small image areas. Of the 42 areas linked to low survival, 34 were in the stroma (supporting tissue), compared with 2 of the 42 areas linked to high survival. The authors concluded that MesoNet effectively predicts overall survival from slide images alone.

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: Development and validation of a deep learning (convolutional neural network) model that predicts overall survival from archived digitized histology slides. The model was trained and tested on a French national database and then checked on an independent US dataset.
• Who was studied: 2,981 pleural mesothelioma patients (one slide each) from multiple French institutions in the MESOPATH/MESOBANK database: 2,300 for training and 681 for testing. A further 56 mesothelioma patients from The Cancer Genome Atlas (TCGA) formed an independent validation set. Study dates are not given.
• Exposure: Not studied. No asbestos or fiber exposure data were analyzed. The article's statement that mesothelioma is associated with asbestos in 80% of cases is cited from other studies.
• Main results:
  • MesoNet ranked patient survival better than the pathologist-assigned histological type (c index). Training cross-validation set: 0.642 versus 0.596. Test set: 0.643 versus 0.598. TCGA set: 0.656 versus 0.590.
  • A simpler method, Meanpool, did slightly better than MesoNet on the MESOBANK training set (0.657 versus 0.642) and test set (0.649 versus 0.643). On the TCGA set, however, Meanpool did worse: 0.581 versus 0.656 for MesoNet.
  • In the test set, the 60 sarcomatoid patients had a median survival of 7.2 months. The 60 patients MesoNet predicted to have the worst prognosis were a mix of types: 34% sarcomatoid and 40% epithelioid. The two groups had comparable outcomes (P = 0.53).
  • In the test set, the 80 grade 1 epithelioid patients had a median survival of 28.1 months. The 80 patients MesoNet predicted to have the best survival were all epithelioid, with a mix of grades, and had similar prognoses (P = 0.3).
  • When test-set patients were split into low, intermediate and high predicted risk, the groups had significantly different outcomes. This held for the whole test set, for epithelioid patients alone, and within each epithelioid grade (grade 1 P = 0.016; grade 2 P = 0.005; grade 3 P = 0.022).
  • Two pathologists reviewed the image tiles most strongly linked to survival. Of the 42 tiles linked to low survival, 34 were in stromal regions, compared with 2 of the 42 tiles linked to high survival.
  • Vacuolization was seen in 27 of the 42 low-survival tiles and in 0 of the 42 high-survival tiles.
  • Some test-set patients had p16 homozygous deletion measured. In these patients MesoNet scored a c index of 0.706, compared with 0.59 for a model based on p16 homozygous deletion only. In patients with p16 expression measured, MesoNet scored 0.726, compared with 0.625 for a model based on p16 expression only.
  • In the TCGA set, only Ki67 expression was significantly correlated with the MesoNet predicted score (correlation of −0.315, P = 0.02).
  • For 167 test-set patients with 2 to 6 biopsies each, the scores for biopsies from the same patient varied significantly less than the scores between different patients.
• The authors' conclusion: The authors concluded that MesoNet, trained on the MESOBANK dataset, effectively predicts overall survival of mesothelioma patients solely from slide images of thoracoscopy or needle biopsies.

Topics

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

Medical terms in this study, explained in our Lexicon: Mesothelioma, Malignant Mesothelioma, Epithelioid, Biphasic Mesothelioma, Sarcomatoid, Overall Survival, Pathology, Inflammation.

Related studies

• Deep-learning based classification distinguishes sarcomatoid malignant mesotheliomas from benign spindle cell mesothelial proliferations — J.R. Naso, A.B. Levine, et al. (2021)
• Asbestos Detection with Fluorescence Microscopy Images and Deep Learning — C. Cai, T. Nishimura, et al. (2021)
• Deep Learning for the Automatic Quantification of Pleural Plaques in Asbestos-Exposed Subjects — I. Benlala, B.D. De Senneville, et al. (2022)
• Prediction of Radiation Pneumonitis With Machine Learning in Stage III Lung Cancer: A Pilot Study — M. Yakar, D. Etiz, et al. (2021)
• Combining a gravitational search algorithm, particle swarm optimization, and fuzzy rules to improve the classification performance of a feed-forward neural network — M.L. Huang, Y.C. Chou (2019)
• The use of a next-generation sequencing-derived machine-learning risk-prediction model (OncoCast-MPM) for malignant pleural mesothelioma: a retrospective study — M.G. Zauderer, A. Martin, et al. (2021)

Browse return to medical research.