Muscle MR Characterization using Unsupervised Learning
Applying Self-Organizing Maps to radiomic data to automate the identification of fatty replacement and edema in STIM1 tubular aggregate myopathy.
Congenital myopathies are highly variable, making standardized diagnosis difficult. This project leverages unsupervised machine learning to identify patterns of muscular degeneration directly from Magnetic Resonance Imaging (MRI) data.
Focusing on a family affected by Tubular Aggregates Myopathy (TAM) due to a STIM1 gene mutation, we extracted comprehensive radiomic features from T1-weighted and STIR MRI sequences. By utilizing Self-Organizing Maps (SOMs), we successfully clustered muscles based on the presence of fatty replacement and edema, demonstrating that an unsupervised AI model can closely replicate expert radiological assessments and track disease progression over time.
The Clinical Challenge
Tubular Aggregates Myopathy (TAM) is a rare congenital condition clinically characterized by muscle weakness, myalgias, and cramps. Tracking the progression of the disease relies heavily on muscle MRI to detect two key pathophysiological changes:
- Fatty Replacement: Evaluated via T1-weighted images, representing the chronic stage of myopathy where muscle tissue is replaced by fibro-adipose tissue.
- Muscle Edema: Evaluated via STIR sequences, indicating acute phases of muscle involvement.
Standardizing these assessments is challenging due to the subjective nature of human radiological evaluation. To assist the diagnostic process, we turned to radiomics—converting standard medical images into minable, quantitative data.
Unsupervised Learning Methodology
Instead of relying on heavily annotated training sets (which are difficult to obtain for rare genetic disorders), we utilized an unsupervised learning approach to see if the data naturally segregated into clinically relevant categories.
- Radiomic Feature Extraction: Using 3DSlicer, we extracted 60 radiomic features (including first-order statistics, GLRLM, and GLCM) from 53 muscles per patient across two MRI assessments conducted 5 years apart ($t_0$ and $t_1$).
- Self-Organizing Maps (SOM): We trained a Kohonen map using the MiniSom Python library. The SOM reduced the high-dimensional radiomic data into a lower-dimensional map, allowing us to group the muscles into clusters based on shared textural characteristics.
- Clustering & Validation: The SOM’s topological weights were grouped using agglomerative clustering. The resulting clusters were then compared against the established Mercuri staging and edema scores assigned by an expert radiologist.
Key Findings
The SOM proved highly effective at discerning muscle states without prior human labeling. It successfully differentiated healthy muscles from those exhibiting pathological MRI alterations. Furthermore, longitudinal analysis between baseline and the 5-year follow-up allowed the SOM to independently identify disease progression, closely mirroring the clinical and radiological worsening observed in the patients.
Research Highlights
Radiomic Pipelines
Transformed qualitative MRI scans into robust quantitative datasets using 3DSlicer for advanced texture analysis.
Self-Organizing Maps
Implemented a Kohonen neural network to achieve unsupervised clustering of healthy, edematous, and fatty muscles.
Longitudinal Tracking
Successfully mapped the natural progression of STIM1 TAM over a 5-year period using purely data-driven methods.
Core Publication
2023
- Muscle magnetic resonance characterization of STIM1 tubular aggregate myopathy using unsupervised learningPLoS ONE, 2023