Deep Learning for Liver Tumor Segmentation
Developed novel U-Net based deep learning models for liver tumor segmentation, achieving state-of-the-art accuracy of 98% using advanced architectures like Squeeze-and-Excitation Networks and Atrous Spatial Pyramid Pooling.
Tech Stack
Results
Achieved state-of-the-art accuracy of 98% for liver tumor segmentation. Co-authored paper published in Elsevier Biomedical Signal Processing and Control.
Key Ideas
- Developed custom data pre-processing pipelines for medical imaging data
- Created custom callbacks, metrics, and loss functions tailored for medical segmentation
- Implemented Squeeze-and-Excitation Networks for channel attention
- Used Atrous Spatial Pyramid Pooling for multi-scale feature extraction
Overview
This project focuses on developing advanced deep learning models for liver tumor segmentation in medical images. The work involved comprehensive data handling, custom architecture design, and achieving state-of-the-art performance.
Architecture
The model is based on U-Net architecture with several enhancements:
- Squeeze-and-Excitation Networks: Channel attention mechanism to focus on relevant features
- Atrous Spatial Pyramid Pooling: Multi-scale feature extraction for better context understanding
- Custom Loss Functions: Tailored for medical segmentation tasks
Technical Contributions
- Comprehensive data pre-processing pipeline for medical imaging
- Custom callbacks and metrics for training monitoring
- Specialized loss functions for medical segmentation
- Modular architecture design for easy experimentation
Results
Achieved state-of-the-art performance:
- 98% accuracy for liver tumor segmentation
- Published in Elsevier Biomedical Signal Processing and Control
- Co-authored research paper detailing methods and results
Impact
This work contributes to the field of medical image analysis, providing accurate and efficient tools for liver tumor detection and segmentation, which can aid in early diagnosis and treatment planning.