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

Area of Science:

• Computational pathology
• Machine learning in oncology
• Gastrointestinal stromal tumor (GIST) research

Background:

• The mitotic index is crucial for GIST grading, but manual counting is time-consuming and prone to errors.
• Existing automated methods are inadequate for GIST spindle cells, highlighting a need for specialized solutions.

Purpose of the Study:

• To develop and validate a machine learning-based automated system for accurate mitosis detection and counting in GIST images.
• To assess the performance of the automated system against manual counts and Ki-67 expression.

Main Methods:

• A GIST image database with 13,965 annotated mitotic cells was created.
• Nuclei segmentation, feature extraction, and selection were performed, followed by training six algorithms.
• A two-step cascaded dual-scale approach using SVM-RBF models at 10x and 40x magnifications was implemented for slide-level counting.

Main Results:

• The SVM-RBF model achieved high performance (F1=0.83 at 10x, F1=0.89 at 40x).
• Automated counts showed moderate correlation with manual counts (r=0.4705) and strong correlation with Ki-67 expression (r=0.6187).
• SHAP analysis confirmed the model aligns with pathologists' criteria.

Conclusions:

• This study presents the first automated framework for GIST mitotic cell detection and counting.
• The findings demonstrate the clinical utility of traditional machine learning in pathology and highlight the model's interpretability.