Title: Performance analysis of 3D brain MRI modalities for brain tumour sub-region segmentation using U-Net architecture

Authors: Anima Kujur; Zahid Raza

Addresses: School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi 110067, India ' School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi 110067, India

Abstract: Magnetic resonance imaging (MRI) is the most widely used technology for brain tumour diagnosis. T1, T1-weighted with and without a contrast agent, fluid-attenuated inversion recovery (FLAIR), and T2-weighted are some of the numerous MRI imaging types available. This work proposes a framework and measures its effectiveness while taking into account the contribution made by each MRI modality in the segmentation of brain tumour sub-regions. The proposed approach has four phases. First phase corresponds to the data collection followed by data pre-processing in the second phase. The third phase entails the implementation of a deep learning model called the U-Net architecture, a form of convolutional neural networks (CNN). The fourth phase evaluates the framework using the popular measures, viz., intersection over union, dice coefficient for each tumour region, sensitivity, and specificity. Simulation study reveals that some MRI modalities contribute comparatively less than the peer MRI modalities to the final performance of the model. Further, it is established that MRIs individually do not perform well but the combination reports an improved performance.

Keywords: magnetic resonance imaging; MRI; medical image segmentation; convolutional neural network; CNN; U-Net; deep learning; fluid-attenuated inversion recovery; FLAIR.

DOI: 10.1504/IJEH.2023.138256

International Journal of Electronic Healthcare, 2023 Vol.13 No.4, pp.311 - 337

Received: 20 Apr 2023
Accepted: 21 Jan 2024
Published online: 30 Apr 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article