Automatic segmentation of the lumbosacral spinal cord

Details

Supervisors

Macq, Benoît ; Jankovski, Aleksandar

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Automatic segmentation of the spinal cord on Medical Resonance (MR) images is crucial in the accurate positioning of spinal cord injuries and the success of the associated treatments. Currently available segmentation algorithms have trouble with the lower part of the spinal cord: the lumbosacral part. This part is often badly segmented or not detected at all. With this work we are trying to close this loophole. Our goal is to create an automatic algorithm dedicated to the lumbosacral segmentation of the spinal cord. This algorithm will be based on an existing one which uses two deep learning models with U-net architecture. The first model is dedicated to the detection of the spinal cord centerline, the second one to the segmentation of images cropped around the detected centerline. By fine-tuning the existing segmentation model it becomes possible to create a U-net model dedicated to the lumbosacral segmentation. The creation of the training set is also part of the work. Manually segmented images are mandatory to have the best possible training. To support this task we have used and improved the CHARP platform developed by the UCLouvain and dedicated to the hosting of medical images and their manual segmentation. To optimize the fine-tuning, we have performed a series of tests to explore the different parameters impacting the training. Theses tests have highlighted the particularities of the U-net model in a fine-tuning process and the importance of the parameters that directly influence the fine-tuning architecture. With the segmentation results of our fine-tuned model, we have identified the two main areas of progress for the fine-tuning and the development of a segmentation algorithm dedicated to the lumbosacral spinal cord. First, we need to fine-tune the centerline detection model similarly to the segmentation model. Very few changes are needed on our current fine-tuning pipeline to support this activity. Second, with only 17 MR images, our training set is too small to ensure a qualitative fine-tuning. Our fine-tuned model provides promising results on some images but is not yet robust enough. Therefore, a larger training set is crucial to continue to develop the model. Our work provides all the bases and tools to complete our initial goal: the automatic creation of segmentation masks for lumbosacral spinal cord.