Cascaded Hybrid Models for Brain Tumor MRI Segmentation under limited dataset conditions

Problem Limited MRI conditions and datasets pose challenges, and the performance remains unknown in testing deep learning models, particularly in brain tumor segmentation. Purpose This study aims to investigate the efficiency, feasibility and effectiveness of integrating a Cascaded U-Sharp network hybrid model using limited MRI conditions and datasets. Methods A Cascaded hybrid model was developed using T1w and FLAIR for the pre-segmentation module with 3D U-Net. The final segmentation module was carried out using pre_results, T1w, FLAIR, T2w, and T1w+C as inputs with residual U-Net. The hybrid model underwent rigorous training and validation using the BraTS2021 datasets, followed by testing with the UPenn-GBM datasets. Moreover, the study conducted a comparative analysis of the proposed model's efficacy with different combinations of inputs using the UPenn-GBM dataset. Results The segmentation accuracy of residual U-Net on the full dataset was evaluated on DICE values of 85.62 (ET), 89.96 (TC), and 92.01 (WT) on the BraTS2021 dataset. Under a limited T2w and FLAIR dataset, the performance was 90.22 (ET), 91.77 (TC), and 92.78 (WT). Conclusion This study proposes a model for brain tumor segmentation. An experiment conducted on comparable and challenging datasets demonstrated better performance using both the full dataset and limited data available.