An improved 3D U-Net-based deep learning system for brain tumor segmentation using multi-modal MRI

• Published in: Multimedia tools and applications Vol. 83; no. 37; pp. 85027 - 85046
• Main Authors: Ali, Saqib, Khurram, Rooha, Rehman, Khalil ur, Yasin, Anaa, Shaukat, Zeeshan, Sakhawat, Zareen, Mujtaba, Ghulam
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.11.2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Brain; Brain cancer; Comparative studies; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Datasets; Deep learning; Encoders-Decoders; Image enhancement; Image segmentation; Machine learning; Magnetic resonance imaging; Medical prognosis; Multimedia Information Systems; Neural networks; Performance evaluation; Special Purpose and Application-Based Systems; Three dimensional models; Track 8: Medical Imaging; Transition layers; Tumors; Brain tumor diagnosis; Deep learning; Multi-modal MRI; Computer-aid diagnosis; Tumor segmentation
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-19406-2

Cancer is a lethal illness that requires an initial stage prognosis to enhance patient survival rate. Accurate brain tumor and their sub-structure segmentation through Magnetic Resonance Images (MRIs) is a tough endeavor. Owing to the heterogeneous tumor areas, automatically segmenting brain tumors has proved to be a critical task even for neural network-based algorithms, some tumor regions remain unidentified due to their small size and the variation in area occupancy among tumor sub-classes. Current progress in the area of neural networks has been employed to enhance the segmentation performance. This study designed an intelligent 3D U-Net encoder-decoder-based system for automatic detection and brain tumor sub-structure segmentation. Our proposed 3D model constitutes neural units (the basic building blocks) followed by transition layer blocks and skip connections. BraTS 2018 and private local datasets are used to evaluate the proposed model which segments the Whole Tumor (WT), Tumor Core (TC), and the Enhancing Tumor (ET). The training accuracy, validation accuracy, dice score, sensitivities, and specificities of WT, CT, and ET zones are computed. The experimental results demonstrate that dice scores are 0.913, 0.874, and 0.801 for the BraTS 2018 dataset. The developed models performance was further evaluated by utilizing the dataset from a local hospital containing 71 subjects. The dice scores of 0.891, 0.834, and 0.776 are achieved by the proposed model on the private dataset. The practicability of the proposed model was assessed by the comparative studies of our model with existing literature.