Advancements in 3D MRI Brain Tumor Segmentation: A Comprehensive Exploration of Fully Convolutional Networks

This work presents a comprehensive study on brain tumor segmentation, utilizing the MICCAI Brain Tumor Segmentation(BraTS) 2016 and 2017 datasets as primary sources for analysis and assessment. The BraTS dataset provides annotated multimodal 3D brain Magnetic Resonance Imaging( MRI), enabling the detection of three tumor classes – Edema (ED), Enhancing Tumor (ET), and Necrotic and Non-Enhancing Tumor (NET/NCR). The study involves the utilization of five distinct 3D Deep Convolutional Networks for analysis and evaluation. Binary Cross Entropy and Dice Loss are investigated as loss functions and three different in-layer normalization techniques are explored using Unet3D to gain a deeper understanding of their impact on performance. Quantitative results reveal that DenseVoxelNet performs poorly due to limited and complex data, while UNet3D consistently outperforms other models. The qualitative analysis demonstrates Unet3D’s superiority over other models, particularly DenseVoxelNet. The findings underscore the significance of skip connections in enhancing the analysis of medical images, revealing a notable 29% increase across multiple evaluation metrics. The classwise evaluation reveals that within the context of NET/NCR segmentation, UNet3D displays a marginal underperformance of 0.7% in terms of Dice Score Coefficient (DSC) and 0.4% in terms of Intersection over Union (IoU). However, UNet3D demonstrates superiority in ED and ET segmentation, exhibiting improvements of 0.4% and 0.9% for DSC and 0.5% and 0.7% for IoU respectively.