Enhancing Auxilliary Cancer Classification Task for Multi-task Breast Ultrasound Diagnosis Network

Breast ultrasound is a non-invasive, economical method that is essential for the diagnosis of cancer. With the advent of deep learning in recent years, numerous CNN-based methods have been thoroughly studied for tasks including tumor localization and cancer classification. While prior single models performed well in both challenges, these approaches had several drawbacks, including long inference times, a need for a GPU, and the need for individual model fine-tuning. Our goal in this work is to create a new end-to-end multi-task architecture that can be used for both segmentation and classification. We obtained exceptional performance and time efficiency with our suggested method, achieving 79.8% and 86.4% in the segmentation challenge in DeepLabV3+ architecture.