Texture Analysis for Muscular Dystrophy Classification in MRI with Improved Class Activation Mapping.

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Neuromuscular Disorders

Background:

• Muscular dystrophies (MD) are rare genetic disorders causing progressive muscle weakness.
• Objective assessment of MD progression is crucial due to limited treatments and outcome measures.
• Fibrofatty tissue replacement in muscles is a key imaging biomarker for MD, particularly Duchenne muscular dystrophy (DMD).

Purpose of the Study:

• To apply deep learning, specifically CNNs, for accurate classification of MD subtypes using MRI data.
• To develop a visualization method for highlighting critical image textures indicative of MD.
• To compare the performance of CNNs against traditional methods for MD assessment.

Main Methods:

• Utilized state-of-the-art Convolutional Neural Networks (CNNs) for MD image classification.
• Employed an Improved Class Activation Mapping (ICAM) technique for visualizing important image regions.
• Evaluated CNN performance on a dataset of dystrophic MRI scans.

Main Results:

• The best CNN model achieved 91.7% classification accuracy for MD subtypes.
• CNNs demonstrated over 40% improvement compared to the traditional Mean Fat Fraction (MFF) criterion.
• ICAM successfully highlighted discriminative texture patterns in specific muscles for DMD and Congenital muscular dystrophy (CMD).

Conclusions:

• CNNs offer a highly accurate and objective method for MD classification and progression monitoring via MRI.
• Deep learning models significantly outperform conventional methods in identifying MD subtypes.
• Advanced visualization techniques like ICAM enhance the interpretability of AI in medical imaging for neuromuscular diseases.