Breast Cancer Segmentation and Classification Using a Hybrid Model
February 2025AI/ML
This project implements a two-stage deep learning approach for breast cancer detection using ultrasound images. It first uses a U-Net model to segment regions of interest, then applies a MobileNetV2-based classifier to categorize images as Normal, Benign, or Malignant. The main goal was to understand how real-world ML pipelines work, from preprocessing to training and evaluation, rather than just achieving high accuracy.
A hybrid deep learning pipeline that segments and classifies breast ultrasound images to assist in medical diagnosis.
Tech Stack
Python
Features
- U-Net based image segmentation for ROI extraction
- MobileNetV2 model for multi-class classification
- End-to-end pipeline from data processing to evaluation
- Handles class imbalance with weighted training
Challenges
- Low model accuracy despite multiple tuning attempts
- Data quality and preprocessing issues
- Difficulty in balancing segmentation and classification performance
- Limited resources for training deep learning models
Learnings
- Learned how hybrid ML pipelines are built
- Gained hands-on experience with CNN models
- Understood challenges of real-world model accuracy
- Improved knowledge of data preprocessing and evaluation