Prediction of Mutations and Outcome in Gastrointestinal Stromal Tumors with Deep Learning: A Multicenter,...

Area of Science:

• Oncology
• Computational Pathology
• Genomics

Background:

• Gastrointestinal stromal tumor (GIST) is the most common gastrointestinal mesenchymal tumor.
• GISTs are driven by KIT and PDGFRA mutations, impacting prognosis and treatment.
• Established risk models for GIST have limited reproducibility; deep learning (DL) on whole-slide images (WSIs) shows promise for molecular classification and prognosis.

Purpose of the Study:

• To develop and validate deep learning models for predicting molecular alterations, treatment sensitivity, and recurrence-free survival (RFS) in GISTs using WSIs.
• To assess the performance of DL models compared to established clinicopathological risk scores.
• To evaluate the prognostic value of DL in different patient subgroups, including those receiving adjuvant therapy.

Main Methods:

• Analysis of 8398 GIST cases from 21 international centers.
• Training DL models on WSIs to predict KIT and PDGFRA mutations, treatment sensitivity (avapritinib, imatinib), and RFS.
• Validation of DL model performance using area under the curve (AUC) and hazard ratios (HR).

Main Results:

• DL models achieved high AUC for predicting mutations: 0.87 for KIT and 0.96 for PDGFRA.
• Specific mutations like KIT exon 11 delinss 557-558 (AUC 0.67) and PDGFRA exon 18 D842V (AUC 0.93) were accurately predicted.
• DL models predicted RFS with HR of 8.44 overall and 4.74 in patients receiving adjuvant therapy, comparable to pathology-based scores.

Conclusions:

• Deep learning on WSIs enables accurate prediction of molecular alterations, treatment sensitivity, and RFS in GIST.
• DL model performance is comparable to established risk scores across international cohorts.
• DL models provide a valuable prognostic tool, particularly in guiding treatment duration decisions for adjuvant therapy.