Automatic Multiple Sclerosis Lesion Tracking Using Unsupervised Machine Learning
Alireza Ahkbardeh1, Ponnada Narayana1, Vladimir Braverman2, Michael Jacobs1, Jason Uwaeze3, John Lincoln4
1Department of Diagnostic and Interventional Imaging, The University of Texas Health Science Center at Houston, 2Department of Computer Science, Rice University, 3Rice University, 4Department of Neurology, McGovern Medical School, UTHealth
Objective:
We used locally linear embedding (LLE) and isometric mapping (Isomap), two Non-Linear Dimensionality Reduction (NLDR) methods, to automatically identify and track enhancing brain lesions on post-contrast MRI in MS.
Background:
Supervised machine learning models are shown to perform image segmentation of different lesions. However, these methods require expert labeling. They are also computationally expensive, require large sample sizes to achieve optimal results, and struggle to detect multi-scale local features like MS lesions. We address these concerns by using unsupervised learning based on NLDR to segment enhancing lesions on brain MRI.
Design/Methods:
The analysis was based on MRIs from 25 deidentified relapsing remitting MS patients. The MRI data included T1 pre-contrast (T1-pre), T1 post-contrast (T1-post), T2 weighted (T2w), Fluid Attenuated Inversion Recovery (FLAIR), and proton density (PD) images. We included T1-pre, T2w, Flair, and PD modalities as a 4-D data into our NLDR methods. As our ground truth, we subtracted T1-post and T1-pre images (subimage). With our 4-D images we used Isomap and LLE to reduce the dimensions to a 2-D image space. We experimented with various neighborhood sizes and found 70 neighbors worked best for both NLDR methods. We created a binary subimage and removed any remaining artifacts not pertaining to MS lesions. The performance of our method was evaluated by using expert identified enhancing lesions on T1-post as ground truth and calculating dice similarity (DS) score.
Results:
We applied our proposed method on 25 subjects and achieved a DS score of 0.80 ± 0.06. All results were compared to a subimage of their respective T1-po st and T1-pre images. The hyperparameters were empirically optimized by looking at different values.
Conclusions:
Our results demonstrate that it is possible to identify the enhancing lesions in brain MRI with unsupervised learning. This opens the possibility of predicting enhancing lesions on unenhanced scans.