Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors Using Segmentation and Feature Extraction

The early detection and accurate classification of breast tumors are crucial for improving treatment outcomes and reducing mortality. This project focuses on enhancing breast tumor analysis using deep learning techniques, specifically leveraging segmentation and feature extraction for classification. The dataset used in this study, Dataset_BUSI, consists of ultrasound images categorized into three classes: benign, malignant, and normal. For segmentation, advanced models like Vision Transformers (ViT) combined with U-Net and Mask R-CNN are employed. These models help in precisely delineating the tumor boundaries within the ultrasound images. Once segmentation is completed, feature extraction is performed using a ResNet-based approach, followed by classification through decision tree, AdaBoost, Naive Bayes, and XGBoost algorithms. These machine learning models are trained to distinguish between benign and malignant tumors effectively. The web-based application developed for this project provides a user-friendly interface, allowing users to upload ultrasound images for segmentation and classification. The application is built using HTML, CSS, JavaScript for the frontend, and Python with Flask for the backend. By integrating deep learning for segmentation and feature extraction, this project aims to assist healthcare professionals in making quicker and more accurate diagnoses.

Keywords: breast tumor classification, segmentation, feature extraction, ultrasound analysis, deep learning, ViT, U-Net, Mask R-CNN, ResNet, XGBoost.