2026-04-17T18:44:05Zhttps://recercat.cat/oai/request

oai:recercat.cat:10230/579452025-12-21T20:46:52Zcom_2072_6col_2072_452954

00925njm 22002777a 4500 dc Gambús i Moreno, Paula author 2023-09-22T17:01:38Z 2023-09-22T17:01:38Z 2023-09-22 Tutors: Dr. Deborah Pareto Onghena, Dr. Gerard Martí Juan. Treball de fi de grau en Biomèdica Multiple sclerosis (MS) is a neurodegenerative disease affecting the central nervous system (CNS), characterized by the destruction of myelin sheaths, that has become the leading cause of disability in young adults. Since this disease does not have a cure, an early diagnosis is crucial to start treatment and slow down its progression. Current diagnostic criteria are based on detecting lesions on magnetic resonance imaging (MRI). Accurate detection of spinal cord (SC) lesions is currently missing. In this context, the aim of this project was to develop an artificial intelligence tool using convolutional neural networks to automatically segment SC lesions from Double Inversion Recovery (DIR) sequences. For this purpose, two different raters manually segmented SC lesions from patients with MS. From these masks, three different networks were obtained using the same hyperparameters: the Rater 1 model was trained with masks segmented by the first rater; the Rater 2 model with segmentations from the rater 2; and the Hybrid one, with masks from both of them. Their performance was evaluated using a test set of 30 patients, with half of them not having SC lesions. To assess the performance of the method, the Dice coefficient between the manual ground truths and the automatic masks was calculated. The three models showed good performance, with a Dice score of around 50%. However, the Rater 2 model had better results since a Dice coefficient of 0.557 ± 0.102 was obtained. Moreover, this algorithm demonstrated a non-overlapping volume percentage of 34.119 ± 49.629 % meaning that it exhibited a tendency to over-segment lesions, leading to false positives. Further analysis and discussions with neuroradiologists determined that removing falsely detected voxels was preferable to missing true SC lesions. Even though the successful results, further improvements to increase its accuracy are necessary for clinical viability. Overall, this research presents a step towards automated SC lesion segmentation in MS using DIR sequences, which could aid to assist radiologists in scanner assessment. Multiple sclerosis Spinal cord lesion Deep learning Convolutional neural network DIR sequence Ground truth Dice coefficient Automatic segmentation of intramedullary multiple sclerosis lesions delimited in DIR sequences with convolutional neural networks