Comparative Analysis of Three U-Net Adaptations for Multiple Sclerosis Lesion Segmentation : A study in collaboration with Sectra

Sammanfattning

Multiple Sclerosis (MS) is a neural disorder that successively degrades the myelin sheath surrounding the neurons. This causes inflammation and scar tissue to develop, called lesions. With the help of MRI, lesions can be detected. At the time of writing, there are no known cures for the disorder and diagnosing the disease early is of importance to minimize the symptoms as the disorder progresses. The lesions are manually segmented by radiologist experts, a time consuming process prone to human factor errors. In recent years, multiple machine learning algorithms have been developed to aid radiologists in this task. In this thesis, the LST-AI, nnU-Net, and 2D U-Net models, trained for lesion segmentation on different data, were evaluated and tested on MRI brain images of MS affected patients. The LST-AI, a pre-trained model developed specifically for MS lesion segmentation, was first tested and compared to the other two self trained models. A total of three datasets were used in training and testing, with exception for the nnU-Net that was only trained and tested on two datasets. The results proved that all models were sensitive to low axial resolution and the lack of proper ground truth for all image sequences had a big impact on their performance. The MRI sequences also had to be used separately and due to missing ground truth for the T1-W and T2-W, their trained models did not perform well. Because of this, no T1-W or T2-W model was tested on other sequences than what they were trained on. The image's sagittal and coronal resolution did not seem to have the same effect as the axial resolution. All three models seemed to perform better on images containing a larger volume of lesions. Overall, given the available data, the nnU-Net models achieved the highest mean Dice scores and IoU, while the self-implemented 2D U-Net model performing the worst.