Classifying presence or absence of calcifications on mammography using generative contribution mapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Mammography is crucial for breast cancer screening.
• Accurate detection of microcalcifications is vital for early diagnosis.
• Explainable AI models can enhance diagnostic tool interpretability.

Purpose of the Study:

• To evaluate the efficacy of Generative Contribution Mapping (GCM) for classifying calcifications on mammograms.
• To compare GCM's performance against EfficientNet classifiers.
• To assess the interpretability of GCM through visualization maps.

Main Methods:

• Utilized 303 full-field digital mammography (FFDM) images from the INbreast database.
• Employed a sliding window method to create calcification and non-calcification patch images.
• Trained, validated, and tested GCM and EfficientNet classifiers on FFDM image patches.

Main Results:

• GCM classifiers achieved high accuracy: 0.92 (CC view) and 0.91 (MLO view).
• GCM demonstrated superior performance with Area Under the ROC Curve: 0.92 (CC) and 0.94 (MLO).
• GCM's visualization maps offered clearer highlighting of regions of interest compared to EfficientNet.

Conclusions:

• GCM effectively classifies the presence or absence of mammographic calcifications.
• GCM provides intuitive and explainable classification grounds via visualization maps.
• GCM shows promise as a reliable tool for aiding in mammographic analysis.