A Machine Learning Approach to Distinguishing between Multiple Sclerosis and Cerebral Small Vessel Disease

Eftekhari, Daniel 1, 2 ; Tyrrell, Pascal 3 ; Martel, Anne 4 ; Moody, Alan 2, 3, 4 ; Aviv, Richard 1, 2, 3, 4

 1. Institute of Biomaterials & Biomedical Engineering, University of Toronto; 2. Sunnybrook Health Sciences Centre, Medical Imaging; 3. Department of Medical Imaging,University of Toronto; 4. Sunnybrook Research Institute, Physical Sciences

Background & Purpose

Multiple sclerosis (MS) and cerebral small vessel disease (SVD) may be difficult to reliably distinguish using only neuroimaging and clinical features. As early diagnosis and treatment may be associated with better long-term outcome in both diseases, and the societal and monetary cost for both diseases is potentially high, particularly for misdiagnosis in MS, we developed an accurate and automatic diagnostic algorithm to distinguish between relapsing-remitting MS (RRMS) and SVD using neuroimaging and clinical features.

Methods

Statistical & machine learning algorithms, including a mixture of t-distributions and a novel spatial heuristic algorithm, were developed for the segmentation of white matter hyperintensities and T1 black holes on FLAIR and pre-contrast T1 W MRI sequences. Combined spatial probability maps of RRMS and SVD lesion masks were subsequently developed using a derivation set of patients. A novel cross entropy image distance metric, which is insensitive to co-registration and anatomical differences between patients, was used to quantify the similarity of new cases to each disease. Bayesian learning algorithms using non-informative priors were trained using these neuroimaging features in combination with clinical features. Model hyperparameters were tuned using Monte Carlo cross validation.

Results

The dataset consists of 39 RRMS (median age 48 (31 – 60), 12 M, 27 F) and 72 SVD (median age 73 (52 – 86), 44 M, 28 F) patients. An intraclass correlation coefficient (two-way random effects, absolute agreement, single measures) of 0.966 (95% CI [0.935, 0.982]) between total lesion volume for the automatic segmentations and RRMS ground truth tracings was obtained. The segmentation algorithms provide the additional benefit of being blind to disease ground truth, and are robust to varying MRI acquisition parameters. The model consisting of both neuroimaging and clinical features obtained 99.3% precision, 97.0% sensitivity, 99.5% specificity, and 98.4% negative predictive value in distinguishing RRMS from SVD on the validation sets; a significant improvement on models developed using neuroimaging and clinical features in isolation. The model misclassified only one case out of 21 on the test set.

Conclusion

As the misdiagnosis rate for MS is 5-10%, and the societal and monetary cost of both diseases is high, this work has real potential for clinical impact. Future directions include using additional predictors thought to distinguish the diseases in the predictive models, and extending the model to multi-disease classification.