A COMPARATIVE EXPERIMENTAL ANALYSIS OF DEEP LEARNING MODELS FOR BRAIN TUMOR IDENTIFICATION AND PREDICTION USING MRI IMAGES

Abstract

Segmentation of brain tumors from magnetic resonance imaging (MRI) remains difficult due to significant variations in tumor regions in terms of size, shape, tissue type and location. In this study, a multimodal deep learning framework named DRA-MSCNet was designed and tested, which included denoising-aware preprocessing, residual learning, multiscale feature extraction and attention-based refinement. The BraTS 2023 Pediatric MRI dataset (99 eligible cases) was used, where 69 cases were used for training, 15 cases for validation, and 15 cases for a locked test set. Whole Tumor (WT), Tumor Core (TC) and Enhancing Tumor (ET) were segmented together from four different MRI modalities: T1n, T1c, T2w and T2-FLAIR. The proposed model achieved Dice scores of 0.7636, 0.7233, and 0.4422 for WT, TC, and ET, respectively, with a mean Dice of 0.6430. Structured pruning was then used to reduce the number of computations. The chosen 10% compressed model was able to reduce the parameters from 2,207,137 to 1,968,491 and the model size from 8.476 MB to 7.567 MB with a mean Dice of 0.6376. There was no significant difference in the original and compressed models (adjusted p = 0.0637) and Mean HD95 improved from 37.8670 mm to 31.4338 mm. Results show that moderate pruning can decrease complexity of computations without sacrificing most of the segmentation performance.