Automatic Segmentation of Intervertebral Discs in Multimodal MRI Using Ensemble U-Net Architecture

Purpose: Intervertebral disc degeneration is a major cause of lower back pain, requiring accurate localization and segmentation for diagnosis and surgical planning.This study aimed to develop a fully automated segmentation method for intervertebral discs using a residual U-Net-based architecture with ensemble learning to improve accuracy across multiple MRI modalities.Methods: A publicly available dataset (IVDM3Seg, MICCAI 2018) containing sagittal MRI scans in four Dixon-based modalities (fat, water, in-phase, out-ofphase) was used.Four separate residual U-Net models with ResNet-34 backbones were trained on each modality using cyclic and layer-specific learning rates for faster convergence.Outputs of these models were combined through a stacked ensemble U-Net for final segmentation.Performance was evaluated using Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and Hausdorff Distance (HD).Results: Individual modality models achieved DSC values between 0.9468 and 0.9754, with MSD ranging from 0.1060 to 0.2208 and HD from 1.0120 to 1.3805 pixels.The ensemble model outperformed individual models, achieving a DSC of 0.9910, MSD of 0.0383, and HD of 0.9855 pixels (p