A better AI-based tool for mesothelioma

Malignant pleural mesotheliomas (MPMs) can be classified into three histological subtypes: epithelioid, biphasic and sarcomatoid. These subtypes, however, do not fully capture the highly variable outcomes of patients with MPM. Now, MesoNet, a new classification tool now described in Nature Medicine, could enable the more accurate prediction of outcomes for these patients.

A deep-learning network was used to assign an overall survival (OS) score based on assessment of 2,048 relevant features in ~10,000 tiles of whole-slide images (WSIs) of tissue specimens from patients with MPM. This machine-learning approach was developed in training (n = 2,300) and testing (n = 681) sets and further validated in WSIs from The Cancer Genome Atlas (TCGA; n = 56).

OS outcomes are typically least favourable for patients with sarcomatoid MPMs. In the test data set, this subgroup comprised 60 patients with a median OS duration of 7.2 months. A similar median OS was observed in 60 patients in this data set predicted by MesoNet to have shorter OS, although this group included a mixture of different histologies (only 34% sarcomatoid). Patients with grade 1 epithelioid MPM typically have the most favourable OS; in the test data set, 80 such patients had a median OS of 28.1 months. The 80 patients in this data set with longer OS predicted by MesoNet had epithelioid MPMs, albeit with a mixture of grades (46% grade 1).

MesoNet outperformed the predictive value of histology-based classification in the training (c index: 0.642 versus 0.596), testing (0.643 versus 0.598) and TCGA (0.656 versus 0.590) sets.

“Our goal was not only to focus on the performance of the model but also on its interpretability, as this is what will really provide value to the medical community,” explains Charles Maussion. The OS scores for all the individual tiles in the original data set were aggregated, and features associated with highest and lowest OS were extracted. Tiles with high OS scores usually had tubular structures and were well vascularized, whereas tiles with low OS scores mainly represented stromal features. “Some of these features were already known but others, such as inflammation, cellular diversity, vacuolization and stromal structures, had not been previously associated with poor prognosis,” comments Maussion. In patients with epithelioid MPM and shorter OS, 39% of tiles displayed these features, which might not have been noticed at diagnosis.

“these features … might not have been noticed at diagnosis”

By enabling a deeper analysis of features related with poor outcomes, this tool has the potential to improve accuracy in the diagnosis of MPM and, importantly, of other malignancies. Prospective validation of this tool is eagerly awaited.


Original article

  1. Courtiol, P. et al. Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nat. Med. 25, 1519–1525 (2019)

    CAS  Article  Google Scholar 

Download references

Author information



Corresponding author

Correspondence to Diana Romero.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Romero, D. A better AI-based tool for mesothelioma. Nat Rev Clin Oncol 16, 722 (2019). https://doi.org/10.1038/s41571-019-0294-1

Download citation


Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing