Lumbar Disc Herniation Detection and Classification in Sagittal, High-Resolution T2-Weighted MRI using Modified Attention U-NET

Lumbar Disc Herniation (LDH) is a prevalent spinal disorder causing severe pain and neurological complications, requiring timely and precise diagnosis. Magnetic Resonance Imaging (MRI) provides detailed visualization of spinal abnormalities, yet manual analysis takes a lot of time and is subject to inter-observer variability. Existing deep learning approaches, such as U-Net, struggle with overfitting and unstable training, limiting their generalization. This study addresses these gaps by proposing a Modified Attention U-Net, integrating batch normalization (BN) after each convolutional layer to stabilize training and dropout (0.3) in the bottleneck layer to mitigate overfitting. Evaluated on T2-weighted sagittal MRI scans of 282 patients (1410 slices), the model achieves 87% classification accuracy, outperforming the baseline U-Net (80%). The validation accuracy curve demonstrates enhanced stability and robustness, unlike the fluctuating performance of the baseline U-Net. These advancements improve feature learning and model generalization, making the proposed method a promising automated tool for LDH detection.