Automated Detection and Segmentation of Breast Lesions in Contrast-Enhanced Mammograms

In this study, we assess an AI-based ensemble of deep learning (DL) architectures designed to auto segment breast lesions in contrast-enhanced mammography (CEM) images. We test this AI system's performance across various densities of breast tissue background enhancement (BPE). Our analysis employed a database of more than 300 CEM images, with a portion reserved for separate validation. We compared the performance for EfficientNetB0, VGG16, and DenseNet121. We have also combined the outputs of the three DL architectures using a weighted average method. This ensemble technique improved lesion segmentation accuracy. The approach achieved a mean Intersection over Union (IoU) score of 0.68 from 10% validation sample. Our result maintained with an IoU of 0.66 (IQR 0.54-0.82) and a dice coefficient of 0.8 (IQR of 0.7-0.9) in an independent testing sample of 21. The results of our study suggest that this AI-based approach has potential to auto segment breast cancer lesion thus to facilitate further AI based assessment such as radiomics assessment. This auto segmentation is a critical step towards complete AI based breast cancer detection and classification.