Towards an interpretable breast cancer detection and diagnosis system

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast cancer screening via mammography is crucial but interpretation can be difficult.
• Computer-aided detection and diagnosis (CAD) systems offer a second opinion but require interpretability for clinical trust.
• Understanding CAD system decisions is vital for effective integration into clinical workflows.

Purpose of the Study:

• To develop an automated and interpretable system for breast cancer detection and diagnosis.
• To enhance the reliability and trustworthiness of CAD systems in mammography.
• To provide a transparent decision-making process for clinicians using AI in breast cancer screening.

Main Methods:

• The system involves pre-processing, unsupervised segmentation, and analysis of textural and shape-based features.
• Feature selection techniques and Decision Tree algorithms are employed for benign/malignant classification.
• Overfitting is addressed using pre-pruning, post-pruning, and Random Forest classifiers for enhanced interpretability.

Main Results:

• The proposed system achieved a maximum accuracy of 95% on the mini-MIAS dataset.
• The system demonstrated 100% precision and 100% specificity in breast cancer detection.
• The system allows users to analyze each step of the detection and diagnosis process.

Conclusions:

• An automated and interpretable system for breast cancer detection and diagnosis has been successfully developed.
• The system provides high diagnostic performance, comparable to expert interpretation.
• The interpretability feature fosters trust and facilitates the adoption of AI in mammography screening.