Multiple Sclerosis Lesion Segmentation with Tiramisu and 2.5D Stacked Slices
Huahong Zhang1, Alessandra M Valcarcel2, Rohit Bakshi3
1Vanderbilt University, Nashville, TN 37235, USA.
Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention
|December 24, 2021
Summary
This study introduces a 2.5D deep learning model for segmenting multiple sclerosis (MS) lesions, improving accuracy by using stacked image slices. The Tiramisu network achieved top performance in a challenge, demonstrating its effectiveness for MS lesion segmentation.
Area of Science:
- Medical image analysis
- Artificial intelligence in medicine
- Neuroimaging
Background:
- Accurate segmentation of multiple sclerosis (MS) lesions is crucial for diagnosis and treatment monitoring.
- Existing methods often struggle with capturing comprehensive contextual information for precise lesion delineation.
Purpose of the Study:
- To develop and evaluate a novel 2.5D fully convolutional densely connected network (Tiramisu) for automated MS lesion segmentation.
- To enhance segmentation accuracy by integrating multi-planar contextual information.
Main Methods:
- A 2.5D approach using stacked slices from three anatomical planes (axial, sagittal, coronal) was implemented.
- The Tiramisu deep learning architecture was employed for lesion segmentation.
- Ablation studies and evaluation on the Longitudinal MS Lesion Segmentation Challenge dataset were conducted.
Main Results:
- The 2.5D Tiramisu network demonstrated competitive performance in MS lesion segmentation.
- Simulated lesion experiments confirmed the benefits of 2.5D patches, stacked data, and the Tiramisu model.
- The L2-loss variant achieved first place on the leaderboard (93.1 overall score), while the focal-loss variant excelled in Dice coefficient (69.3%) and lesion-wise true positive rate (60.2%).
Conclusions:
- The proposed 2.5D Tiramisu network offers a significant advancement in automated multiple sclerosis lesion segmentation.
- Integrating multi-planar contextual information effectively improves segmentation accuracy.
- The method shows strong potential for clinical application in MS patient management.
