Visual mapping of muscle MRI fatty replacement patterns in genetic myopathies using dimensionality reduction

Muscle MRI is a complementary diagnostic tool in genetic myopathies; however, its interpretation remains challenging because numerous muscles must be assessed simultaneously, and disease-specific patterns overlap. This study evaluated whether dimensionality reduction techniques (DRT) can transform complex muscle MRI fatty replacement data into meaningful low-dimensional visual maps of patient similarity. We analyzed a large multicenter dataset comprising 975 patients with genetically confirmed diagnoses across ten myopathies. Mercuri scores of pelvic and lower-limb muscles were used as input. Principal Component Analysis (PCA), ISOMAP, t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) were applied to generate two-dimensional representations of muscle fatty replacement. The quality of these low-dimensional maps was quantitatively assessed by measuring the extent to which they preserved disease-specific organization, using unsupervised clustering as a proxy of consistency. Gaussian Mixture Models (GMM) were subsequently applied to assess whether the low-dimensional maps retained sufficient information to support disease discrimination. UMAP and t-SNE outperformed PCA and ISOMAP. Furthermore, UMAP produced more coherent and better-separated disease groupings than t-SNE, reflected by a higher V-measure metric (0.415 vs 0.403), and achieved superior top-3 diagnostic accuracy when combined with GMM (87%vs 81%). Overall, dimensionality reduction provides a framework for visualizing muscle MRI patterns similarity across neuromuscular diseases supporting pattern recognition.