Improving Multiple Sclerosis Classification: A Novel Machine Learning Approach Using Multimodal Retinal Imaging Data

Theme: Neuro-ophthalmology
What: Neuro-ophthalmology
Part of: Neuro-ophthalmology I / Neuro-ophtalmolgie I
When: 6/1/2024, 11:15 AM - 12:45 PM
Where: Room | Salle 714 B

Abstract

Purpose: Multiple sclerosis (MS) is a progressive neurodegenerative disorder affecting approximately 2.8 million individuals globally. The optic nerve has been proposed as a fifth anatomic compartment to determine dissemination in space for MS diagnosis. To this end, machine learning analysis of imaging data from color fundus photography (CFP) and optical coherence tomography (OCT) may be a promising source of diagnostic biomarkers. Although there is tremendous potential to develop machine learning models for MS classification, the high dimensionality of multimodal CFP and OCT image data poses significant challenges for model training, especially for smaller datasets. As a result, most existing MS classification approaches predominantly utilize low-dimension preprocessed OCT tabular features for training. However, this reduction in data dimensionality may lead to important diagnostic information being lost during preprocessing, ultimately decreasing the clinical performance of MS classification models. The purpose of this study was to develop a novel machine learning architecture for MS classification using multimodal CFP and OCT image data that could be effectively trained on small datasets. Furthermore, we compared the performance of the proposed model to a MS classifier trained solely on preprocessed OCT tabular features.

Study Design: Retrospective cross-sectional study

Methods: We developed a deep learning architecture based on RETFound, a foundation model pretrained on 1.6 million retinal images, to perform MS classification using multimodal CFP and OCT image data. The performance was compared against an XGBoost model trained for MS classification on 24 preprocessed OCT tabular features. Training and validation were performed using a balanced dataset constructed from 124 UK Biobank participants (62 with MS and 62 healthy control). Five-fold cross validation was used to evaluate the performance of the models.

Results: The model trained using multimodal CFP and OCT image data achieved an AUC (area under the receiver operating curve) of 0.83±0.04, which was significantly larger (p<0.05) than the AUC of the model trained using OCT tabular features, which was 0.77±0.02.

Conclusions: The improved performance of the proposed model suggests that multimodal CFP and OCT imaging data likely contain pertinent diagnostic information for MS classification which is not captured by preprocessed OCT tabular features. Consequently, training machine learning models using multimodal CFP and OCT image data may be advisable to achieve optimal MS classification performance. Incorporating the developed methodologies into clinical practice has potential to elevate diagnostic precision for patients with MS and other neuro-ophthalmologic conditions where diagnostic biomarkers are discernible from retinal imaging data.

Presenter(s)

Presenting Author: Christopher Nielsen

Additional Author(s):

Fiona Costello, University of Calgary

Matthias Wilms, University of Calgary

Nils Forkert, University of Calgary

---