Clinically-inspired automatic classification of ovarian carcinoma subtypes

Area of Science:

• Computational pathology
• Digital image analysis
• Machine learning in oncology

Background:

• Ovarian carcinoma subtypes are distinct entities with significant prognostic and therapeutic implications.
• Pathologist histotyping is reproducible but can be challenging, necessitating immunohistochemistry and consultations.
• An automated framework is needed to enhance diagnostic accuracy and streamline pathologists' workflow.

Purpose of the Study:

• To develop and validate an automated framework for ovarian carcinoma subtype classification.
• To improve the accuracy and reproducibility of ovarian carcinoma diagnosis.
• To assist pathologists in their diagnostic procedures.

Main Methods:

• Analysis of histopathology images at two magnification levels.
• Extraction of color, texture, and shape descriptors using image processing.
• A machine learning pipeline including dissimilarity matrix, dimensionality reduction, feature selection, and SVM classification.

Main Results:

• The system achieved 95.0% multiclass classification accuracy on unseen histopathology images.
• The classifier's confusion matrix aligns with clinical observations, particularly for endometrioid and serous carcinomas.
• Validation was performed on a dataset of eighty high-resolution images.

Conclusions:

• Ovarian carcinoma diagnosis is challenging due to intrinsic class imbalance among subtypes.
• Automated analysis of ovarian carcinoma subtypes can serve as a valuable second opinion for clinicians.
• The proposed system demonstrates potential to aid in the diagnostic procedure.