Download PDFOpen PDF in browserEnhancing U-Net for Breast Tumor Segmentation with Automatic Data AugmentationEasyChair Preprint 1251513 pages•Date: March 16, 2024AbstractSegmentation plays a crucial role in computer-aided diagnosis (CAD) for the early detection and diagnosis of diseases. This involves identifying and locating the regions of interest (ROI). Despite recent successes in computer vision technology, challenges persist due to the limited accessibility of medical data, resulting in models that may not generalize well to unseen data. This paper proposes an approach utilizing automatic data augmentation techniques for enhancing U-Net performance in breast tumor segmentation. We employ data augmentation in both the training and prediction phases. During training, augmentation increases the diversity and quantity of training data. In the prediction phase, Test Time Augmentation (TTA) is used to aggregate segmentation outputs from several versions of the original input, generated by the augmentation policy. Experimental results on the BrEaST dataset of 256 ultrasound breast tumor images and corresponding masks demonstrate the effectiveness of our proposed method. It improves U-Net's Dice Similarity Coefficient (DSC) from 0.7078 to 0.7541, concurrently mitigating the risk of overfitting and ensuring robust generalizability to unseen data. This research contributes to improving segmentation outcomes on limited datasets and promises to enhance the robustness of CAD systems in medical image analysis. Keyphrases: Automatic Data Augmentation, Breast tumor segmentation, Medical Image Analysis, U-Net, test-time augmentation
|