Deep Learning Predicts Mutations and Outcomes in Gastrointestinal Stromal Tumors
Arianna Bonetti1, Van-Linh Le2,3,4, Zunamys I Carrero1
1Else Kroener Fresenius Center for Digital Health, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany.
Cancer Research
|June 8, 2026
Summary
Deep learning models analyzing whole-slide images accurately predict mutations and treatment sensitivity in gastrointestinal stromal tumors (GIST). These models also predict recurrence-free survival, matching established risk scores for GIST patients.
Area of Science:
- Oncology
- Computational Pathology
- Genomics
Background:
- Gastrointestinal stromal tumors (GIST) are the most common mesenchymal tumors of the GI tract.
- Mutations in KIT and PDGFRA are key drivers, influencing prognosis and treatment response.
- Current risk models have limited reproducibility, highlighting the need for advanced tools.
Purpose of the Study:
- To develop and validate deep learning (DL) models for molecular classification and prognostic assessment of GIST using whole-slide images (WSIs).
- To predict specific mutations, treatment sensitivity to tyrosine kinase inhibitors (TKIs), and recurrence-free survival (RFS) in GIST.
- To compare the performance of DL models against established clinicopathological risk scores.
Main Methods:
- Analysis of 8398 GIST cases from 21 international centers, including molecular and clinical follow-up data.
- Training DL models on WSIs to predict KIT and PDGFRA mutational status, including specific variants.
- Evaluating DL model performance in predicting sensitivity to avapritinib and imatinib, and predicting RFS.
Main Results:
- DL models achieved high AUCs for predicting KIT (0.87) and PDGFRA (0.96) mutations, including specific subtypes.
- Accurate prediction of avapritinib (0.84) and imatinib (0.81) sensitivity.
- DL models effectively predicted RFS (HR 8.44 overall), with performance comparable to pathology-based scores.
Conclusions:
- Deep learning applied to WSIs is a powerful tool for predicting molecular alterations, treatment sensitivity, and RFS in GIST.
- DL models demonstrate comparable or superior prognostic performance to traditional risk scores across diverse international cohorts.
- This approach provides a foundation for developing multimodal predictors for improved GIST management.
