Brain Tumor Segmentation on MRI BraTS Data using convolutional neural networks: experimental study and comparisonDragičević, Irena (Avtor) Žibert, Janez (Mentor) Fošnarič, Miha (Recenzent) master's thesesradiologic technologyconvolutional neural networksU-NetBraTSglioma segmentationmultimodal segmentationmedical image segmentationmachine learningIntroduction: The research on medical image segmentation is now-a-days largely driven by convolutional neural networks. These networks perform state-of-the-art results for brain tumor segmentation. A modification of convolutional neural networks called U-Net has gained significant popularity among researchers. Purpose: The aim of this work is to develop a 3D U-Net model in Python. The same architecture is trained using different sequences and the segmentation results of each model are evaluated and discussed. Methods: The 3D U-Net is developed in Python using the BraTS 2020 dataset. The purpose of the 3D U-Net model is to segment input brain MRI images of patients with glioma into several classes and subregions while also implementing metrics to assess its effectiveness. The dataset was split to a training and a validation set and segmentations from U-Net models are validated by calculating the Intersection over Union and the Dice similarity coefficient for classes and tumor subregions. Results: The Intersection over Union and Dice similarity coefficient show variations in the segmentation performance depending on the sequences used to train the U-Net model. Overall, the whole tumor is best segmented by the T2 and FLAIR combined U-Net, the tumor core by the U-Net trained on T1CE and FLAIR and the active tumor by the U-Net trained on T1CE and T1. The T1CE and T1 combination has shown the same performance for necrosis/non-enhancing tumor segmentation as the T1CE and FLAIR combination, while the T2 and FLAIR combination was inadequate for the task. The same can be said for the segmentation of the enhancing tumor while the T1CE and FLAIR-trained model was the most useful for edema segmentation. Discussion and conclusion: The results have shown that different sequences play an important role in the training process of a U-Net and significantly affect the validation.. Overall, the U-Net model trained on T1CE and FLAIR has shown the best performance as it has accomplished a very effective segmentation of all classes and tumor subregions, with little to no drawbacks when compared to other models. The clinical implementation of U-Net models could bring improvement to the hospital workflow and greatly benefit the patient.[I. Dragičević]20232023-06-29 07:45:41Magistrsko delo/naloga147286sl