Automatic Segmentation of Muscle Tissue and Inter-muscular Fat in Thigh and Calf MRI Images

Background: MRI of thigh muscles is a commonly used technique to estimate the fat infiltration into muscles, thus an accurate segmentation of muscle and fat is required. The grey-level intensity in each imaged pixel is formed from a combination of two signals-fat and water, caused by the infiltration of subcutaneous fat into the muscle tissue. This leads to unreliable segmentation if conventional segmentation methods are used. Hence, to address this issue, we propose a deep-learning-based method for muscle segmentation.

Methods:

Data 17 patients, 5 slices for each, the T₂ and PD maps were constructed.

Muscle segmentation a deep convolutional-neural-network (CNN) called U-net was trained on 14 patients for the muscle segmentation. In order to increase the number of training images, augmentation was applied to the training set.

Results: We trained the CNN on pre-processed T2 and PD maps and reached 94.5% dice on test-set consisting of 3 patients. The ability of the CNN to reach challenging results was explored by feeding the CNN with T2 and PD maps without pre-processing and the raw data in another experiment, the performances were 95% and 95.2% respectively. A 6-fold cross-validation was used for assessing the generalization of the model and to estimate the segmentation accuracy. The averaged result was 96.1% dice.

Conclusion: We suggested a deep learning-based method to accurately segment muscles in mild, moderate and severe fat infiltration. We also demonstrated the reliability of the proposed method when intensity inhomogeneity artifacts exist in MRI images and confirmed the generalization of this method.