Deep learning models predict the risk of relapse and the mutational profile in GIST

An algorithm developed by deep learning from digitalised haematoxylin and eosin-stained whole tumour slide images outperformed classical AFIP/Miettinen relapse risk prediction in patients with gastrointestinal stromal tumours (GIST). A second algorithm predicted mutations with high accuracy.

GIST, the most frequent mesenchymal tumour of the GI tract, shows a variable clinical behaviour ranging from benign to malignant. Risk assessment according to the AFIP/Miettinen classification (high, intermediate, or low risk of relapse) and mutational profiling are major tools for patient management. The AFIP/Miettinen classification includes parameters such as the size of the tumour, localisation, and mitotic count [1]. However, the AFIP/Miettinen classification comes with subjectivity (mitotic count), costly mutational profiling and is time-consuming, and not yet available in all countries (or centres). Therefore, Dr Raul Perret (Institute Bergonié, France) and colleagues evaluated the efficacy of 2 deep learning models: 1 to predict relapse-free survival in GIST patients and 1 to predict mutational profiles [2].

Both models were based on histology, i.e. digitised haematoxylin and eosin-stained whole tumour slide images. The relapse-predicting model was trained using whole tumour slide images from 305 patients (Institut Bergonié) and validated using slide images from 286 patients (Léon Bérard Centre). Both cohorts had a similar distribution of GIST types (localisation, TKI treatment). The mutation profile-predicting model was trained using images from 1,233 patients (Institut Bergonié) and validated on images from 238 patients (Léon Bérard Centre).

The algorithm for relapse prediction outperformed prediction based on AFIP/Miettinen classification (C-index 0.81 vs 0.76). Combining deep learning with tumour location and tumour size (Deep Miettinen), further improved C-index to 0.83. Deep Miettinen was able to stratify patients into subgroups at high or low risk for relapse-free survival. Also, the algorithm was able to split patients, characterised as ‘high risk for relapse’ according to classical AFIP/Miettinen, into 2 additional groups: high versus low risk. Likewise, the algorithm was able to split classical AFIP/Miettinen ‘intermediate risk’ patients into high-risk and low-risk groups.

The algorithm for the prediction of the presence of mutations reached an area under the receiver operating characteristic curve (AUC) for predicting KIT mutations of 0.80 in the training cohort and 0.85 in the validation cohort. AUC for predicting PDGFRA mutations was 0.92 in both cohorts. More specific, AUC for predicting PDGFRA exon 18^D824V mutation was 0.87 in both cohorts and predicting KIT exon 11^{del 557–558} mutation was 0.69 in the training cohort and 0.76 in the validation cohort.

“Our results strongly suggest that implementing deep learning with digitised whole slide images may represent a reproducible way to improve tailoring therapy and precision medicine for patients with GIST. However, further validation of the models is needed,” concluded Dr Perret.

1. Miettinen M, Lasota J. Semin Diagn Pathol. 2006;23(2):70–83.
2. Italiano A, et al. Deep learning predicts patients’ outcome and mutations from H&E slides in gastrointestinal stromal tumor (GIST). Abstract 1484O, ESMO Congress 2022, 09–13 September, Paris, France.

Copyright ©2022 Medicom Medical Publishers



Posted on 16 November 202227 March 2024