Enhancing Breast Cancer Diagnosis With Attention Branch Network and Thermographic Imaging

Abstract

The high mortality rate among breast cancer patients in developing regions is primarily due to the lack of affordable access to breast screening systems for the detection of abnormalities. Thermographic breast screening aided by machine learning-based decision support systems has shown promising results. We present an interpretable computer-assisted diagnostic system that enhances clinical inference by visual identification of regions of interest in thermographic images. A CNN feature extractor with an Attention Branch Network (ABN) is developed for binary classification of thermographic images. We trained and validated our model on a newly created Amrita Breast Thermogram (ABT) dataset consisting of 331 participants. The model performance compared against standard clinical mammogram results demonstrated an F1 score of 98.88% (precision: 97.78%, recall: 100%, accuracy: 98.15%) after sample weighting. The model was also tested on another publicly available dataset, DMR-IR, wherein the ABN-DCN model demonstrated comparable performance (accuracy: 95%). Test results showed that incorporating the ABN along with sample weighting enhanced the performance of the baseline DarkNet19 CNN model by 6%. The proposed DarkNet19-integrated ABN decision support system offers diagnostic interpretability besides top-tier performance.