An interpretable machine learning framework for automated mitosis detection in gastrointestinal stromal tumors

Abstract

The mitotic index is a critical indicator in grading gastrointestinal stromal tumors (GIST). Conventional microscopy-based mitosis counting is labor-intensive and exhibits interobserver variability, necessitating automation. However, existing models have proven unsuitable for GIST spindle cells. To address this limitation, we developed a machine learning method for automated mitosis detection in GIST. A GIST image database, annotated with 13,965 mitotic cells, was first established. Following nuclei segmentation, feature extraction, and feature selection, six different algorithms were employed to train mitosis detection models on images at both 10 × and 40 × magnification levels. The Radial Basis Function Support Vector Machine (SVM-RBF) achieved the best performance under both magnifications (10 ×: F1 = 0.83; 40 ×: F1 = 0.89). Slide-level mitosis counting was then performed via a two-step cascaded dual-scale approach, in which the 10 × model first identified Regions of Interest (ROIs), followed by precise detection and counting using the 40 × model. Slide-level validation showed a moderate correlation between automated and manual mitosis counts (r = 0.4705) and a strong correlation between the automated counts and Ki-67 expression (r = 0.6187). SHAP interpretability analysis confirmed that the model's decision-making basis closely aligned with pathologists' diagnostic criteria, including nuclear membrane disintegration, chromatin condensation, and chromosomal alignment. In summary, this study establishes the first automated framework for mitotic cell detection and counting in GIST. It underscores the clinical potential of traditional machine learning for targeted pathological applications and demonstrates favorable interpretability.