A novel multi-atlas segmentation approach under the semi-supervised learning framework: Application to knee cartilage...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Multi-atlas based segmentation methods are effective but computationally expensive due to exhaustive patch libraries and voxel-wise labeling.
• Existing techniques face challenges with memory requirements and long execution times.

Purpose of the Study:

• To propose a novel two-stage multi-atlas approach under the Semi-Supervised Learning (SSL) framework to address the computational cost of segmentation.
• To improve segmentation accuracy and efficiency in medical imaging applications.

Main Methods:

• A two-stage multi-atlas segmentation approach utilizing SSL, exploiting spectral content and incorporating unlabeled target data.
• Integration of sparse reconstructions, HOG feature descriptors, and graph-based label propagation.
• Stage-1 focuses on global data sampling and labeling; Stage-2 addresses out-of-sample data.

Main Results:

• Superior segmentation performance (DSC:88.89%, Precision:89.86%, Recall:88.12%) compared to patch-based methods on 76 subjects from the Osteoarthritis Initiative (OAI) repository.
• Over 70% reduction in average execution time compared to patch-based methods.
• Comparable performance to state-of-the-art methods in 3-class and 5-class knee cartilage segmentation settings.

Conclusions:

• The proposed semi-supervised learning method offers a computationally efficient and accurate alternative for medical image segmentation.
• The approach demonstrates significant improvements in performance and speed, outperforming existing patch-based and deep learning methods.