Deep learning-based automated liver contouring using a small sample of radiotherapy planning computed tomography
N Arjmandi1, M Momennezhad2, S Arastouei3
1Department of Medical Physics, School of Medicine, Mashhad University of Medical Sciences, Mashhad, Iran; Student research committee, Mashhad University of medical sciences, Mashhad, Iran.
Radiography (London, England : 1995)
|August 23, 2024
Summary
Deep learning accurately segments livers using limited data for radiotherapy planning. Modified U-Net models show strong generalizability on unseen datasets, demonstrating feasibility with small training sets.
Area of Science:
- Medical imaging and artificial intelligence
- Radiotherapy planning and image segmentation
Background:
- Limited data availability for deep learning in medical imaging
- Need for efficient liver contouring in radiotherapy
Purpose of the Study:
- Investigate minimum data requirements for deep learning-based liver contouring
- Assess feasibility of automated liver segmentation with limited data
Main Methods:
- Preprocessing of radiotherapy planning CT images (denoising, windowing)
- Segmentation using modified Attention U-Net and Residual U-Net networks
- Evaluation of model performance on internal test sets and two unseen external datasets
Main Results:
- Modified Residual U-Net achieved 97.62% Dice Similarity Coefficient (DSC) with 62 training cases
- Average Hausdorff Distances were 0.57 mm (Residual U-Net) and 0.71 mm (Attention U-Net)
- High DSC scores (95.16%-95.82%) on external datasets indicate strong generalizability
Conclusions:
- Deep learning models can achieve accurate liver segmentation with small training datasets
- Proposed method demonstrates robustness and generalizability on unseen data
- Potential for automated liver contouring in radiotherapy planning
