AI and imaging advances revolutionize multiple sclerosis diagnosis

By Dr. Liji Thomas, MDReviewed by Benedette Cuffari, M.Sc.Jul 21 2024

Recent advances in imaging technology and artificial intelligence (AI), including machine learning (ML), have facilitated the use of optical coherence tomography (OCT), which is often combined with infrared reflectance scanning laser ophthalmoscopy (IR-SLO), to automate the diagnosis and classification of multiple sclerosis (MS). A recent study in the journal Translational Vision Science & Technology determines the potential independent use of IR-SLO images to automate the diagnosis of MS.

Study: SLO-Net: enhancing multiple sclerosis diagnosis beyond optical coherence tomography using infrared reflectance scanning laser ophthalmoscopy images. Image Credit: Ground Picture / Shutterstock.com

Automating the diagnosis of MS

MS is an autoimmune disease that arises due to the demyelination and breakdown of the central nervous system (CNS) and neural axons. Currently, the diagnosis of MS depends on the presence of clinical symptoms, laboratory tests, such as oligoclonal bands due to the presence of abnormal immunoglobulin antibodies in the cerebrospinal fluid (CSF), and magnetic resonance imaging (MRI).

OCT reflects early changes in the thickness of the retinal layers, which may indicate progressive atrophy of the retinal nerve fibers (RNF) in the RNF layers (RNFL) and optic nerve due to preceding acute inflammation. Thinner retinal layers typically correlate with worse visual outcomes.

OCT, which often correlates with MRI findings, is useful for identifying MS subtypes and is associated with both physical and cognitive disability. It is also a biomarker of the progress of neurodegeneration, which allows clinicians to monitor advancing disability and the efficacy of treatments.

Recently, AI has been used to automate MS diagnosis by integrating MRI, serum, CSF, and OCT results. Despite their widespread use in retinal imaging, it remains unclear how IR-SLO or monochromatic fundus imaging may function as diagnostic tools.

IR-SLO compensates for eye movements and increases OCT B-scan alignment, thereby reducing noise and ensuring minimal variability on follow-up imaging. IR-SLO images show some structures that are poorly seen on fundus imaging due to different wavelengths being used in these tests.

Compared to healthy patients, OCT angiography of MS patients exhibits significant variations in the retinal vessel density. To date, these differences have not been reported by human evaluation of IR-SLO images.

Currently, ML-based MS diagnostic tools have only been trained on OCT thickness measurements. The current study examines the possibility of improving the discriminative ability of ML models by applying deep learning (DL) to IR-SLO images and OCT to identify subtle vascular or structural changes present in MS.    

To our knowledge, this is a pioneering study incorporating IR-SLO into automated diagnosis of MS…among the few studies that have applied DL to retinal imaging data for detecting MS.”

About the study

Both OCT and supplementary IR-SLO images were obtained from 32 individuals with MS and 70 healthy people as controls. Two separate databases were created using multiple convolutional neural networks (CNNs) trained on both OCT and IR-SLO data. This included 132 IR-SLO and 124 OCT images from healthy controls and 133 IR-SLO and 60 OCT images from MS patients.

This produced a bimodal model capable of achieving superior performance compared to each model trained with IR-SLO images or OCT thickness map alone.

Training, validation, and testing sets were created stringently to avoid data leakage. The testing set included 27 and 29 IR-SLO and OCT images from healthy controls, respectively, and 24 IR-SLO and 14 OCT images from MS patients.

The bimodal prototype was associated with 92% accuracy, 95% sensitivity, and 92% specificity, respectively. The test showed 97% discriminative ability, as demonstrated by the area under the receiver operating characteristic (AUROC) curve.

The area under the precision-recall curve (AUPRC) was 97%, thus indicating an excellent balance between false positives and false negatives.

The broad applicability of this model was examined by substituting the internal dataset with an external dataset for testing. This led to an accuracy of 85%, 97% sensitivity, and 85% specificity, as well as 99.7% for both AUROC and AUPRC curves.

Conclusions

While the model trained only with IR-SLO did not perform as well as that trained only with OCT thickness data, using OCT maps and IR-SLO in a merged model led to 3% higher accuracy and sensitivity in diagnosing MS than OCT-based models alone. This is achieved by using complex models with many more parameters to ensure the extraction of multiple features from both images to effectively discriminate between MS and healthy patients.

We showed that a hybrid CNN receiving input data from both modalities can detect MS with astonishing ACCs near 100%”, “showcasing the potential of utilizing IR-SLO as a complementary tool alongside OCT.”

An AI-based system can identify minor changes that are significant for optic nerve disease that may escape human observation. In fact, optic neuritis characteristic of MS is typically observed by physicians only after the loss of over half of the RNFL.

OCT is much less expensive and invasive than MRI. With the same device capable of providing IR-SLO images, similar ML models could be incorporated into routine clinical practice for the automated diagnosis of MS in the future.

In the future, larger and more diverse validation studies are needed, as the current study used a dataset obtained from a single center with limited samples.