Master thesis presentation by Kokchun Giang, MPBME

On Monday, 12 June, Kokchun Giang, MPBME, will present his master thesis with the title “Automatic segmentation of knee muscles in MRI data using deep learning convolutional neural network”. Examiner: Fredrik Kahl

Monday, June 12 at 11.00

Where? Lunnerummet (room 3311), Hörsalsvägen 11, 3rd floor, Chalmers


Osteoarthritis in the knee is a disease that breaks down the cartilage, which causes tremendous pain among the patients. In many cases this require surgical intervention such as total knee arthroplasty. One way to help increasing consistency and accuracy of prosthesis placement is through image guided surgery, which require automatic image segmentation of relevant anatomical structures around the knee, in particular the muscles. During recent years, with stronger and cheaper GPUs, convolutional neural networks have excelled at image recognitions and segmentation tasks. The aim of this thesis is to create an automatic segmentation algorithm based on 3D convolutional neural networks to segment muscles in 3D MRI data.

During training, voxels from 14 patients were sampled and a 3D patch were created around each sampled voxel. These patches were then used as input to the network to train weights and bias parameters with purpose to correctly classify the center sampled voxel. The output from the network is a probability of the voxel having foreground class (muscle) or background class (non-muscle).

Further, in the training process, 2 patients were used as validation data of the network. Hyperparameters are changed to improve validation accuracy so that the network can generalize better to data that it has not been trained on. Finally, to reconstruct the segmentation of the test patients, a forward pass was done for each voxel of each test patient. The final segmentations were obtained by thresholding the network outputs, applying morphological opening and closing. Mean Dice coefficient for the final segmentation was 0.8786 for foreground voxels and 0.9724 for background voxels.