649P Comparing unsupervised AI techniques for visualizing MRI fat infiltration patterns in muscular dystrophies

• Published in: Neuromuscular disorders : NMD 2024-10, Vol.43, p.104441, Article 104441.175
• Main Authors: Pizarro-Galleguillos, B., Goméz-Andrés, D., Tobar, F., Andia, M., Díaz-Manera, J., Verdú-Díaz, J., Rojas, L. Suazo, Jara, J. Díaz, Bevilacqua, J.
• Format: Article
• Language: English
• ISSN: 0960-8966
• DOI: 10.1016/j.nmd.2024.07.184

Muscular dystrophies (MD) are a group of genetic disorders caused by mutations in genes involved in muscular structure and function. They are characterized by muscular weakness and dystrophic histopathological changes. MRI is a supportive diagnostic tool due to its sensitivity in detecting muscle fat infiltration patterns (FIP). Different FIPs have been associated with specific MDs; for this reason, MRI has gained a role in their diagnosis. Heatmaps have been used to represent MRI's FIP. However, it is difficult to compare the FIP of patients with different diseases using heatmaps because of their high dimensionality. For example, a lower limb MRI heatmap can have up to 70 muscles. We hypothesized that dimensionality reduction techniques (DRT) could effectively represent MRI FIP in a low-dimensional space, allowing easier visualization and comparison of patients. Four DRTs were compared: PCA, ISOMAP, t-SNE, and UMAP. An open MRI's FIP database of 975 patients with a genetically confirmed diagnosis of 10 different MDs was used. The database consists of lower limb MRI muscle fat infiltration scores semi-quantitatively graded through the Mercuri scale from T1w images. To quantify the performance of the DRTs, a K-means clustering algorithm was run 20 times with different seeds over the dimensionality-reduced coordinates. Six metrics of clustering results performance were compared (Misclassification fraction, Homogeneity, Completeness, V-measure, Adjusted Rand Index, Adjusted mutual information, and Silhouette coefficient). These results were compared using unpaired t-tests and corrected for multiple comparisons. UMAP significantly outperformed the other techniques, with a small but significant difference with t-SNE in all metrics except for homogeneity. t-SNE outperformed PCA and ISOMAP in all metrics, and similarly, ISOMAP outperformed PCA. In this way, we showed that DRTs are suitable tools that can facilitate the visualization of the MRI's FIP of patients with MDs. Although further research is needed to validate these results in diverse patient populations and clinical settings, DRTs, especially t-SNE and uMAP, are promising tools that may aid in the visualization of fat infiltration patterns in patients with DMs, which may facilitate the diagnosis of these disorders in the clinical setting.