Unsupervised muscle region extraction by fuzzy decision based saliency feature integration on thigh MRI for 3D modeling

Abstract

As the first and important step of modeling to explore biomechanical dynamics of the human body, the regions of interest, such as muscles, bones, nerves, and etc., should be extracted from slices of MRI or CT data. Fast and automatic region segmentation would speed up an online process of building physically based models. In this work, a new automatic segmentation model was proposed that utilizes both saliency-on and saliency-off features with simple morphological operations and binary fuzzy decision based fusion. The new model was tested on a dataset of 160 images including two different T1 and one T2 thigh parts. 3D models of the thigh parts were successfully generated with a high segmentation performance, achieving over 90% F-measure values based on 2D comparison and low RMS values of Hausdorff distance based on 3D volumetric comparison. That is adequate for 3D model research since smoothing on manual or automatic segmented images is also applied during 3D construction of the thigh muscle structure. Experimental results showed the possibility of providing automatic muscle segmentation using saliency features for modeling 3D human musculoskeletal system.