ATTENTION-GUIDED U-NET WITH GAN DISCRIMINATOR AND CNN-BASED SHAPE CLASSIFIER FOR BREAST TUMOR DETECTION

Breast cancer causes the highest number of cancer- related deaths among women worldwide, making accurate classification and analysis of breast tumors in ultrasound imaging crucial for early diagnosis and treatment planning. In this work, we propose using a two-stage deep learning framework in segmentation and classification of breast ultrasound tumors analysis. In Stage I, we implement an Attention U- Net as a generator in a PatchGAN adversarial framework to perform segmentation of the tumor as accurately as possible. The model in Stage I is developed with adversarial loss and Tversky loss that allows us to overcome the class imbalance, and improve accuracy on the boundaries of the object. In Stage II, the tumor masks from Stage I are resized and classified as normal,benign,malignant using a conventional neural network (TumorNet).The masks are also classified as an irregular, lobular, oval and round tumors using a custom convolutional neural network (DeepShapeNet). The proposed framework was trained and validated on the Breast Ultrasound Images (BUSI) public dataset. The results demonstrate the segmentation model significant Dice scores and binary accuracy, and the classification model showed a strong ability to simply observe the shape and classify correctly based confusion matrix. As an integrated approach, the proposed framework offers a reliable and interpretable pipeline for providing automated assessment of breast tumors in clinical ultrasound imaging. 1