Multiple sclerosis lesion segmentation in 2D T2-FLAIR brain magnetic resonance imaging

Abstract

This study aims to address the gap in automatic segmentation of Multiple Sclerosis lesions within 2D T2-FLAIR magnetic resonance imaging. The project was developed using a small basis dataset available on the Mendeley website, consisting of original images and corresponding segmentations executed by professional radiologists and neurologists. Brain tissue extraction and data augmentation techniques were used to address the challenge posed by the lack of prior documentation, representing a significant advancement in preprocessing steps for this particular MRI modality. During the brain extraction procedure, FSL from Oxford University was employed, namely BET (brain extraction tool), which was originally created for T1 modality images. This required necessary additional preprocessing of image pixels to resemble T1 images, including contrast adjustment. To ensure that image augmentation is well performed on both original image and segmentation, the transformers package from the Python environment was used. The augmentation consisted of simple procedures such as rotations, flips and noise addition with Gaussian blur. The investigation has successfully developed and validated an automatic segmentation model for T2-FLAIR MRI images. The model achieved a remarkable Dice coefficient of almost 0.6 in a transformer architecture with 12 layers and attention heads, indicating substantial agreement with ground truth annotations. Despite using a relatively small dataset, the results demonstrate the feasibility of the approach in clinical settings. The performance is commendable for its initial scope. However, to improve the accuracy and generalizability of the model, further data enrichment is necessary. The success in extracting brain tissue and augmenting data, along with the encouraging segmentation outcomes, demonstrates the potential for more comprehensive studies. Future efforts should focus on expanding the dataset and refining the model to enhance the diagnostic and monitoring capabilities of Magnetic Resonance Imaging in Multiple Sclerosis management. This study presents a new approach to medical image processing and sets a fundamental precedent for future investigations into the automated segmentation of Multiple Sclerosis lesions in T2-FLAIR modality.

Date of Award	4 Jun 2024
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	António Silva Ferreira (Supervisor)