Automatic segmentation of the human thigh muscles in magnetic resonance imaging

Shaubari, Ezak Fadzrin Ahmad (2018) Automatic segmentation of the human thigh muscles in magnetic resonance imaging. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (23MB) | Preview

Abstract

Advances in magnetic resonance imaging (MRI) and analysis techniques have improved diagnosis and patient treatment pathways. Typically, image analysis requires substantial technical and medical expertise and MR images can su↵er from artefacts, echo and intensity inhomogeneity due to gradient pulse eddy currents and inherent e↵ects of pulse radiation on MRI radio frequency (RF) coils that complicates the analysis. Processing and analysing serial sections of MRI scans to measure tissue volume is an additional challenge as the shapes and the borders between neighbouring tissues change significantly by anatomical location. Medical imaging solutions are needed to avoid laborious manual segmentation of specified regions of interest (ROI) and operator errors. The work set out in this thesis has addressed this challenge with a specific focus on skeletal muscle segmentation of the thigh. The aim was to develop an MRI segmentation framework for the quadriceps muscles, femur and bone marrow. Four contributions of this research include: (1) the development of a semi-automatic segmentation framework for a single transverse-plane image; (2) automatic segmentation of a single transverseplane image; (3) the automatic segmentation of multiple contiguous transverse-plane images from a full MRI thigh scan; and (4) the use of deep learning for MRI thigh quadriceps segmentation. Novel image processing, statistical analysis and machine learning algorithms were developed for all solutions and they were compared against current gold-standard manual segmentation. Frameworks (1) and (3) require minimal input from the user to delineate the muscle border. Overall, the frameworks in (1), (2) and (3) o↵er very good output performance, with respective framework’s mean segmentation accuracy by JSI and processing time of: (1) 0.95 and 17 sec; (2) 0.85 and 22 sec; and (3) 0.93 and 3 sec. For the framework in (4), the ImageNet trained model was customized by replacing the fully-connected layers in its architecture to convolutional layers (hence the name of Fully Convolutional Network (FCN)) and the pre-trained model was transferred for the ROI segmentation task. With the implementation of post-processing for image filtering and morphology to the segmented ROI, we have successfully accomplished a new benchmark for thigh MRI analysis. The mean accuracy and processing time with this framework are 0.9502 (by JSI ) and 0.117 sec per image, respectively.