Brain tumor image segmentation using model average ensembling of deep networks

• Published in: International journal of system assurance engineering and management 2024-08, Vol.15 (8), p.3915-3925
• Main Authors: Mishra, Ajey Shakti, Acharya, Upendra Kumar, Srivastava, Akanksha, Modi, Aashi Rohit, Kumar, Sandeep
• Format: Article
• Language: English
• Subjects: Engineering; Engineering Economics; Logistics; Marketing; Organization; Original Article; Quality Control; Reliability; Safety and Risk
• ISSN: 0975-6809; 0976-4348
• DOI: 10.1007/s13198-024-02392-x

In the biomedical field, identification of brain tumors along with their location, regions of spreading, and speed of extension are of utmost importance to decide the treatment for Brain Tumors. Automated segmentation plays a major role in detection because manual extraction of the brain tumor sub-regions from MRI volume is monotonous, error-prone, and intricate. Deep learning significantly contributed to outperforming these issues since it is aware of their complexity. Therefore, a technique for the automated segmentation of MRI brain pictures has been developed using model average ensembling of deep networks such 3D CNN and U-Net architectures. 3D CNN and U-Net architecture have made remarkable progress on the task of segmentation of brain tumors. Due to their reliability, ensembling of these models have been opted to have a model with greater reliability. The novelty of this paper is to build a robust segmentation technique by model average ensembling of 3D CNN and U-Net models for abnormality identification by improving the image quality using preprocessing methods. The model includes the testing set BraTS-19 as its input dataset. After performing a lot of experiments, it has been observed that the obtained dice scores by the proposed model for TC (Tumor Core), WT (Whole Tumor), and ET (Enhancing Tumor) are 0.9603, 0.9201, and 0.9237 respectively. The obtained dice scores from the ensembling technique are better than existing techniques. The demonstrated results show the supremacy of the proposed method with an overall accuracy greater than 96%.