Novel Deep Learning Technique Detects White Matter Lesions Faster in Early MS

New machine learning techniques can help neurologists detect active white matter lesions during the early stages of multiple sclerosis (MS) that may otherwise be difficult to capture using current scanning methods, according to study results presented at the Americas Committee for Treatment and Research in Multiple Sclerosis (ACTRIMS) Forum 2023, held in San Diego, California, from February 23 to 25.

Researchers in Australia conducted a clinical study to assess whether machine learning and artificial intelligence could detect and more accurately diagnose active white matter lesions in the early stages of MS compared to existing imaging.

The inventors developed an automated algorithm architecture based on GoogleNet trends that used 6 convolutional layers containing 256 filters in the first layer and incrementally decreasing in filter size from 128 down to 8 filters as the last layer. This is known as a convolutional neural network (CNN) as opposed to fully connected (FC) layers, which current scanning methods use.

The researchers also capitalized on data augmentation and preprocessing techniques, including skull stripping, image denoising, and intensity normalization. These preprocessing techniques allowed for cleaner images and greater pixel diversity prior to processing the images with the main algorithm.

They obtained 171 magnetic resonance imaging (MRI) scans from 3 different tesla MRI scanners. They used 61 scans to train the algorithm and the remaining 110 scans to test the validity of the algorithm.

Next, the researchers evaluated the precision, sensitivity, and dice similarity coefficient (DSC), the latter of which is a metric of spatial overlap and reproducibility, of this algorithm using CNN as compared with existing diagnostic tools.

The researchers discovered that this new automated algorithm increased overall average accuracy of detection of active white matter lesions from 88% to 92%. They also confirmed the robustness of the new algorithm when it was applied to different scanners with varying deep learning parameters, suggestive of its effective performance during real-world application.

Additionally, the new algorithm was faster, analyzing between 200 to 300 image slices in an average time of 2 minutes, indicating that use of this more efficient concept may save neurologists time in obtaining a more precise MS diagnosis in the early stages of the disease.  

“This novel fully automated algorithm … was faster … compared to … existing methods, leading to more accurate and robust performance,” the researchers noted. “This automated algorithm will help neurologists to diagnose MS at an early stage when treatment is most efficacious,” they concluded.