Comparing lesion segmentation methods in multiple sclerosis: Input from one manually delineated subject is sufficient
M M Weeda1, I Brouwer1, M L de Vos1
1Department of Radiology and Nuclear Medicine, MS Center Amsterdam, Amsterdam Neuroscience, Amsterdam UMC, location VUmc, De Boelelaan 1118, 1081 HV Amsterdam PO box 7057, Amsterdam 1007 MB, The Netherlands.
Neuroimage. Clinical
|November 18, 2019
Summary
Retraining nicMSlesions with a single multiple sclerosis (MS) patient dataset significantly improves lesion segmentation accuracy. This deep learning approach offers better volumetric and spatial agreement than untrained methods for MS lesion analysis.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence in Medicine
Background:
- Accurate segmentation of multiple sclerosis (MS) lesions is crucial for monitoring disease progression, including lesion load and brain atrophy.
- Convolutional Neural Network (CNN) strategies, like nicMSlesions, are increasingly preferred in MS lesion segmentation challenges.
- Existing CNN models may be trained on homogenous data, potentially limiting their generalizability to diverse datasets.
Purpose of the Study:
- To evaluate the performance of the nicMSlesions CNN model on an independent dataset.
- To compare nicMSlesions against manual segmentation and other automated methods.
- To determine the suitability of nicMSlesions for multi-center multiple sclerosis studies.
Main Methods:
- Manual lesion segmentation was performed on 3D FLAIR MRI scans from 14 MS patients.
- Five categories of automated segmentation methods were compared: unsupervised/untrained, supervised/untrained, supervised/untrained with threshold adjustment, supervised/trained (leave-one-out), and supervised/trained (single-subject).
- Volumetric accuracy was assessed using the intra-class correlation coefficient (ICC), and spatial accuracy using Dice's similarity index (SI).
Main Results:
- Supervised and trained methods (nicMSlesions, BIANCA) achieved the best volumetric and spatial agreement with manual segmentation (ICC > 0.968, median SI > 0.643).
- The unsupervised, untrained method (LesionTOADS) performed the worst (ICC = 0.140, median SI = 0.444).
- Single-subject retraining of nicMSlesions showed variable but adequate performance, outperforming untrained methods, especially when initial lesion volumes were low.
Conclusions:
- Retraining the nicMSlesions deep learning CNN with data from a single multiple sclerosis patient is sufficient for adequate lesion segmentation.
- This single-subject trained approach demonstrates superior volumetric and spatial agreement compared to untrained methods like LesionTOADS and LST-LPA.
- The findings support the adaptability of nicMSlesions for larger, multi-center studies requiring robust MS lesion segmentation.
