Joris M. van Sabben

DL accurately predicts key molecular mutations in GIST from histopathology. A, AUC with 95% CIs for DL-based prediction of selected molecular mutations in the internal validation cohort (pink) and the external validation cohort (green). B, Representative top-ranked predictive image tiles for three different mutations (PDGFRA exon 18 D842V, KIT exon 11 with two or more codons deleted, and KIT exon 9), as identified by the model’s attention score. Green arrows, lymphocytes for KIT exon 9 and PDGFRA exon 18 D842V; red arrows, vacuolization of cells for PDGFRA exon 18 D842V; blue arrows, mitoses for KIT exon 11 ≥ 2 codons/del–ins. C, AUC with 95% CIs for the same model shown in A when deployed on an external validation cohort consisting exclusively of biopsy specimens; results are reported for all selected mutations. D, Bar plot detailing the sensitivity (pink) and specificity (light pink) for each mutation. E, Heatmaps generated via gradient-weighted class activation mapping (Grad-CAM) illustrate the predictive importance of individual tissue tiles for each mutation: red, high prediction score; blue, low prediction score. In the case of well-characterized mutations such as PDGFRA exon 18 D842V and KIT exon 9, the model focused on tumor regions, in line with reported morphologic correlates. In other cases (PDGFRA exon 18 other mutations), mucosal regions were also emphasized, likely due to vacuolated cells resembling PDGFRA-mutated morphology. Even for more specific variants, such as KIT exon 11 deletions or insertions, the model continued to highlight relevant tumor areas.

DL stratifies recurrence risk and complements clinical prognostic models in GIST. A–C, Analyses in the overall C1 cohort: A, correlations between the DL score and the mitotic index assessed by the Spearman rank test; B, Kaplan–Meier curves for RFS according to the DL score (binarized at the median value of the C1 training dataset). C, C-indices with 95% CIs for five prognostic models: mitotic model (“mitosis”), simple DL model (“DL score”), pathologic MJ model (“Path. MJ”), continuous MJ model (“Cont. MJ”), and deep MJ model (“Deep MJ”). D–F, Corresponding analyses in the C2 subcohort (patients from C1 who did not receive adjuvant TKI therapy). G–I, Corresponding analyses in the C3 subcohort (patients from C1 who received adjuvant TKI therapy). J, Top representative tiles. Tiles 1, 2, and 3 from high-risk tumors showed high cellularity with epithelioid cells, nuclear atypia, and hyperchromasia. Tile 1 showed numerous mitoses (arrows). Tiles 4, 5, and 6 from low-risk tumors showed low cellularity with spindle cells and fibrosis, no atypia and no hyperchromasia. ***, P < 0.001.

Supplementary Tables from Deep Learning Predicts Mutations and Outcomes in Gastrointestinal Stromal Tumors

Sensitivity, specificity, and F1-score for molecular mutation prediction.

Sensitivity and specificity for treatment-sensitivity prediction.

Kaplan-Meier curves for recurrence-free survival (RFS) and overall survival (OS) in the subgroup of C2 and C3 patients with high risk score according to pathological Miettienen-Joensuu scoring system and Imatinib sensitive mutations depending on the deep Miettinen-Joensuu models. Kaplan-Meier curves for RFS (A) and OS (B) depending on the deep Miettinen-Joensuu model employing the C2 DL Score in this subgroup of the C2 subcohort. Kaplan-Meier curves for RFS (C) and OS (D) depending on the deep Miettinen-Joensuu model employing the C3 DL Score in this subgroup of the C3 subcohort. *: p < 0.05; ***: p < 0.001. Tests are log-rank tests.

DL predicts treatment sensitivity and supports clinical decision-making in GIST. A, Standard clinical workflow from initial diagnosis to treatment selection in patients with GIST. B, Proposed integration of the DL model into the diagnostic–therapeutic pathway, enabling early triage of patients for further molecular analysis or tailored treatment selection. C, AUC with 95% CI for DL-based prediction of treatment sensitivity, reported for the internal validation cohort (pink), external validation cohort (green), and an external biopsy-only validation cohort (light green). D, Confusion matrices for avapritinib and imatinib sensitivity, shown at the optimal threshold by Youden’s index (green) and at the threshold yielding the highest F1 score (pink).

Statistical pipeline for recurrence-free survival (RFS) analysis. (A) Data partitioning. Boxes with light dashed contours corresponded to the patients with available DL Scores while boxes with solid thick contours to the patient included in the survival analysis. (B) Methodological approach performed in C1, C2 and C3 to obtain deep learning (DL) scores and to develop survival models in the Training cohorts using the Cox proportional hazard (CPH) algorithm and to compare them. Other abbreviations: ext.val: external validation, int. val: internal validation, M.J.: Miettinen-Joensuu. *Covariables were: age, sex, adjuvant TKI therapy and mutational status for C1; and age, sex and mutational status for C2 and C3.

Prognostic performance of mitotic index, DL score, and MJ risk models in cohorts C1 to C3.