Development of a DL Model Trained to Predict MS Wins Top Award
RSNA Kuo York Chynn Neuroradiology Research Award presented during RSNA 2025
A study on the development of a multimodal deep learning (DL) model trained on structural and diffusion MRI (dMRI) to predict multiple sclerosis (MS) won the 2025 Kuo York Chynn Neuroradiology Research Award for the top neuroradiology research paper presented during the RSNA annual meeting.
MS is the most common central nervous system demyelinating disease in the U.S. Despite this, its diagnosis is challenging. “There is no definite biomarker for MS,” said Jiajian Ma, MD, a PhD student in data science at New York University in New York City. “Its clinical and imaging presentations are very heterogeneous and may overlap with mimics like migraines and cerebral small-vessel disease.”
According to Dr. Ma, white matter lesions and diffuse injury in normal-appearing white matter (NAWM) are two key pathological changes of MS. While lesions are identifiable in conventional structural MRI (sMRI), more subtle NAWM injuries are often hidden in sMRI and require dMRI methods to be detected.
However, Dr. Ma said dMRI is not incorporated in current MS diagnostic criteria, nor consistently acquired in routine clinical workflows. MS diagnosis therefore relies on lesion-centric imaging biomarkers, which can sometimes result in overdiagnosis when patients have MS mimics.
“I believe NAWM injury could serve as a complementary biomarker,” Dr. Ma said. “It can help to differentiate MS from vascular-related white-matter lesions, which usually lack widespread NAWM abnormalities.”
Receiver operating characteristic (ROC) curves under FLAIR-only inference on the external test set, comparing models trained with sMRI+dMRI versus sMRI-only. (A) Before lesion masking: Models trained with sMRI+dMRI (red, purple) achieve higher ROC-AUC than sMRI-only models (blue, green), indicating that incorporating dMRI during training improves generalization when applied to sMRI-based inference. (B) After lesion masking (illustrated in C–D): sMRI+dMRI trained models (red, purple) retain superior performance despite removal of lesion signals, suggesting greater sensitivity to normal-appearing white matter (NAWM) abnormalities.
AI Sharpens Detection of MS
To develop a DL model capable of capturing hidden NAWM information in sMRI to predict MS, Dr. Ma and colleagues trained it on a multimodal dataset comprising 8,118 structural and 3,621 dMRI scans.
The model’s performance was evaluated on both an internal test set (n=838, from NYU Langone Health) and an external test set (n=1,756, from 14 public datasets), each including a diverse spectrum of MS mimics such as white-matter lesions caused by migraine, cerebrovascular disease and non-MS demyelinating disorders.
“We incorporated dMRI during training to guide the model toward learning NAWM-related features of MS, leading to stronger sMRI-based representations for distinguishing MS from its mimics, without requiring additional dMRI scans at inference,” Dr. Ma explained.
When tested using sMRI alone, the model demonstrated excellent performance on both the internal (AUC = 0.965, 95% CI [0.942–0.985]) and external test sets (AUC=0.976, 95% CI [0.968–0.983]).
In an ongoing reader study, the model has also shown promising trends toward higher sensitivity and specificity than established MS diagnostic biomarkers, including
- Dissemination in space, such as with MS lesions appearing in different areas of the brain or spinal cord.
- Dissemination in time, meaning new MS lesions forming at different points in time.
- The central vein sign, which refers to the appearance of a small vein running through the center of an MS lesion, helping distinguish it from other conditions.
- Paramagnetic rim lesions, recognized as MS lesions with a dark outer ring visible on advanced MRI.
Incorporating dMRI during training substantially improved the model’s generalizability on external datasets. When using fluid-attenuated inversion recovery (FLAIR) images for inference, the AUC increased from 0.921 to 0.970.
Even after lesion removal on FLAIR, the model trained with both dMRI and sMRI maintained superior performance (AUC = 0.867 vs. 0.766 for sMRI-only training).
“We were surprised that including diffusion MRI during training made the model much better at making inferences from data across different centers,” Dr. Ma said. “Even when we digitally ‘filled in’ brain lesions, the model still performed okay, suggesting that brain regions outside visible lesions may also provide important clues for MS diagnosis.”
“We were surprised that including diffusion MRI during training made the model much better at making inferences from data across different centers. Even when we digitally ‘filled in’ brain lesions, the model still performed okay, suggesting that brain regions outside visible lesions may also provide important clues for MS diagnosis.”
— JIAJIAN MA, MD
A Step Toward More Individualized Diagnosis, Care
MS diagnoses are usually made by neurologists, with radiologists supplying imaging evidence that aids the neurologist’s assessment. Dr. Ma and his colleagues designed their model to facilitate this process by providing quantitative evidence beyond conventional imaging biomarkers. Such evidence included an MS probability score, an associated confidence estimate and a saliency map.
“Our current framework is specifically designed for multiple sclerosis, but it can be readily adapted for related tasks,” Dr. Ma said. “For other neurological disorders, many research groups are exploring similar deep-learning approaches using brain MRI. Together, these efforts could gradually transform how we use imaging to understand and manage complex brain diseases.”
While the initial results are promising, Dr. Ma acknowledged that there remain key challenges to implementing DL-based tools in clinical practice.
“Whether these tools can truly influence future diagnostic criteria will depend on how well they demonstrate real-world clinical value,” he said. “Our study is primarily retrospective, and prospective deployment will be needed to confirm model performance and clinical impact in real-world settings.”
Funded with a donation by internationally renowned neuroradiologist and longtime RSNA member Kuo York Chynn, MD, the RSNA Kuo York Chynn Neuroradiology Research Award provides $3,000 to the author of the top neuroradiology research paper presented at the RSNA annual meeting.
“I’m optimistic,” Dr. Ma concluded. “As evidence accumulates—for example, through prospective studies or clinical trials—AI or AI-derived biomarkers could serve as decision-support tools to assist in MS diagnosis.”
For More Information
Learn more about the Kuo York Chynn Neuroradiology Research Award.
Read RSNA News stories on the previous years’ winners of the research award: