The following is a summary of “Improving lesion detection in mammograms by leveraging a Cycle-GAN-based lesion remover,” published in the February 2024 issue of Oncology by Lee et al.

In response to the challenge posed by the diverse appearance of breast lesions and normal structures on mammograms, researchers aimed to enhance the performance of computer-aided detection algorithms for breast cancer. Their strategy involved developing a Lesion Highlighter (LH) leveraging a Cycle-GAN-based Lesion Remover (LR). Utilizing 10,310 screening mammograms from 4,832 women, comprising 4,942 recalled lesions (BI-RADS 0) and 5,368 normal results (BI-RADS 1), investigators divided the dataset into Train:Validate: Test folds (0.64:0.16:0.2). Employing a Cycle-GAN, the study group created two Generative Adversarial Networks (GANs) to transfer the style of lesions to normal breast tissue, referred to as LR. 

The study group highlighted the lesions by color-fusing the mammograms after applying the LR. Three deep networks (ResNet18, DenseNet201, EfficientNetV2, and Vision Transformer) were trained on highlighted mammograms (Highlighted), original mammograms (Baseline), and a combination of both (Combined). ROC analysis on the test set demonstrated that the Combined version of all networks achieved improved AUCs (0.963 to 0.974) compared to their Baseline counterparts (0.914 to 0.967), with statistical significance (p-value < 0.001). These results underscore the effectiveness of the Cycle-GAN-based LR in enhancing lesion conspicuity, thereby improving the overall performance of the detection algorithm. 

This approach holds promise for advancing the accuracy of breast cancer screening in mammography.

 Source: breast-cancer-research.biomedcentral.com/articles/10.1186/s13058-024-01777-x