Breast CEM Image Analysis: Tumor Recognition, Segmentation, and Classification - A Workflow Evaluation

We trained our model on our internal dataset and evaluated its performance on a public dataset to demonstrate the effectiveness of our workflow for breast CEM image recognition and segmentation of tumors, followed by classification of benign and malignant tumors. Although these cases were labeled as cancer cases, the majority of them only had one breast affected by breast cancer, both in our internal dataset and the public dataset. Therefore, the dataset we used includes both malignant and non-malignant images, with a roughly balanced distribution of positive and negative samples. Our dataset consists of x cases, with each case having 4 contrast-enhanced images (LCC, RCC, LMLO, RMLO) and 2 low-energy images (CC and MLO views of the affected side). Hence, there are a total of x images, with x malignant and x non-malignant images. We employed random cropping and random translation as data augmentation techniques. We horizontally flipped all images of the right breast and resized all images to 224\u00d7224 before feeding them into the model.