Comparative Analysis of Deep Learning Models for Brain Tumor Segmentation in MRI Scans Using BraTS and Experimental Datasets

Motivation: Precise brain tumor segmentation from MRI is essential for accurate tumor volume calculation and assessing treatment effectiveness, but traditional methods are labor-intensive and inaccurate, driving exploration of deep learning approaches. Goal(s): To evaluate the performance of six UNet variants on BraTS and private-dataset, assessing adaptability and accuracy to identify the best model for further research. Approach: Comparative analysis of nnU-Net, U-Net++, ELU-Net, Attention-UNet, and UNETR, using Dice and Jaccard metrics. Transfer-learning was applied to mitigate dataset-specific performance drops. Results: nnU-Net demonstrated the best performance on both datasets, reaching an 89% Dice score on BraTS and 84.5% on the private-dataset after transfer-learning. Impact: It highlights potential of deep-learning models, particularly nnU-Net to improve accuracy and efficiency in brain-tumor segmentation, reducing reliance on labor-intensive methods like RANO and iRANO. Addressing dataset-specific limitations through transfer-learning, the findings aids in consistent tumor-volume assessment, enhancing treatment monitoring.