Gaussian mixture modelling enables valid intramuscular fat quantification on T₂- and intermediate-weighted shoulder...

To evaluate the accuracy and reliability of automated Gaussian Mixture Modelling (GMM) and K-means clustering for intramuscular fat (IMF) quantification on T₂-weighted (T₂W) and intermediate-weighted (IW) shoulder MRI, using two-point Dixon as the reference standard.

This retrospective study collected MRI datasets of patients who underwent both two-point Dixon with T₂W and/or IW shoulder MRI. GMM and K-means were applied to quantify IMF of the supraspinatus, subscapularis, infraspinatus, teres minor, and deltoid. Accuracy and reliability relative to Dixon were evaluated using Bland-Altman analysis and intraclass correlation coefficients (ICC_2,1). Raw IMF values from GMM, K-means and Dixon were compared using repeated-measures ANOVA with post-hoc Dunnett's test (∝=0.05).

Eighty patients (mean age, 58 years ± 15 [standard deviation]; 41 males) were evaluated. On T₂W MRI, GMM outperformed K-means with a maximum mean bias of -3.07% and excellent reliability for supraspinatus, subscapularis, infraspinatus and deltoid (ICC_2,1 ≥ 0.90). K-means produced poor to moderate reliability. On IW MRI, GMM mean bias did not exceed -2.81% and showed excellent reliability for infraspinatus (ICC_2,1 = 0.92) with good reliability for other muscles. K-means showed higher bias (-4.36%) and lower reliability (ICC_2,1 ≤ 0.81) than GMM. GMM was the only model that did not significantly differ from Dixon IMF values on T₂W (p ≥ 0.074) and IW (p ≥ 0.243) MRI.

IMF quantification using GMM thresholding achieves better agreement with Dixon IMF values than K-Means on T₂W and IW MRI.