The digital eye for mammography: deep transfer learning and model ensemble based open-source toolkit for mass detection and classification

Abstract

Breast cancer stands as a prevalent malignancy affecting women globally, and a screening method, mammography, boasts reliability for early diagnosis. Nevertheless, interpretive errors during population screening may result in false negatives and positives. To address this, Computer-Aided Detection systems rooted in deep learning have emerged, aiming to reduce both false positive and negative predictions. This study introduces an open-source toolkit called The Digital Eye for Mammography (DEM) and addressing limitations in mammography screening for mass detection and classification. The DEM comprises 11 state-of-the-art object detection architectures and uses a meticulously labeled dataset. It serves as a transfer learning source, and provides ensemble of models from diverse deep-learning architectures, resulting in a more robust solution. Experiments conducted on widely-used datasets indicate that the DEM outperforms existing transfer learning sources by significant margins in terms of true positive rate (TPR). According to the experimental results, the DEM serves as a better transfer learning source for mass detection in pathology-proven InBreast and CBIS-DDSM datasets, presenting improvements 12% and 5% in TPR performance at 0.1 false positive per image (FPPI), respectively. Compared to literature, the DEM achieves lower FPPI values while maintaining higher sensitivity, indicating its potential usage as a transfer learning source. By employing ensemble strategies, the DEM produces more reliable outcomes in our KETEM dataset, reducing FPPI by 49% for BI-RADS 1-2 (Breast Imaging Reporting and Data System) and 46% for BI-RADS 4-5 compared to the best individual model while preserving TPR values. The DEM's results suggest its ability to attain better performance without requiring complex model hyperparameters optimization. The GitHub repository of the DEM project is publicly available on: https://github.com/ddobvyz/digitaleye-mammography.