This is the final group project developed along with @MirkoMorello and @andypalmi for the Signal and Imaging Acquisition and Modelling in Healthcare course at Unimib. Canva presentation
Breast cancer is the most common cancer in women globally, with early detection playing a crucial role in patient survival. This project aims to assist radiologists by developing an AI-powered tool that can automatically segment breast lesions from ultrasound images and classify them as benign or malignant.
The project is composed of two main tasks:
-
Lesion Segmentation: Semantic segmentation was tackled using state-of-the-art deep learning architectures:
- DeepLabV3+ (with ResNet34, ResNet50, and Xception65 backbones)
- UNet++ (with ResNet34 backbone)
These models were trained on 647 ultrasound images paired with pixel-wise lesion masks. Images were augmented during training to improve generalization. For patients with multiple lesions, segmentation masks were merged via bitwise OR.
-
Lesion Classification: Each segmented lesion was analyzed through radiomic features (101 per lesion), extracted and used to train three different classifiers:
- Support Vector Machine
- Random Forest
- Feed-Forward Neural Network
The classification focused on correctly identifying malignant lesions (high sensitivity), with evaluation based on metrics such as F1, sensitivity, specificity, and accuracy.
See the slides for visualizations, architecture diagrams, and more technical details.