Deep Learning-Based Vision Approach For Early Detection and Classification Of Breast Cancer

Breast cancer remains one of the major reasons behind the death rates among women all over the world. Proper and early diagnosis is crucial to proper treatment outcome. In this work, we present a novel deep learning-based technique designed to automatically detect and segment breast tumors in ultrasound imaging. Our approach integrates statistical texture priors, Gabor feature extraction, and a hybrid CNN-LSTM architecture trained on stage-specific tumor features. Additionally, super pixel segmentation enhances lesion localization, facilitating effective biopsy guidance. This method is inspired by recent advancements, notably the STAN and ESTAN models by Shareef (2023) and HoVer-Trans by Moetal. (2023), which address smalltumor detection, anatomical context, and noise challenges in ultrasound imaging. Our system achieves high accuracy, efficient lesion boundary segmentation, and robustness in low-contrast conditions, making it clinically valuable for early stage cancer detection.