A Dual-Branch Self-Enhancing Network for Lumbar Disc Herniation Lesion Segmentation and Classification

Abstract

Lumbar Disc Herniation (LDH) is characterized by a complex anatomical structure and significant morphological variability on magnetic resonance imaging (MRI). These characteristics, in turn, present significant challenges for automated diagnosis. To address this, this paper proposes a novel multi-task deep learning architecture, the Dual-Branch Self-Enhancing Network (DBSE-Net), designed to simultaneously perform key structure segmentation and disc classification on LDH MRI, thus improving the efficiency of computer-aided diagnosis. DBSE-Net first constructs a Frequency Self-Enhancing Encoder (FSE Encoder), composed of stacked Frequency Self-Enhancing Block (FSE Block). Each block uses Frequency-Enhanced Convolution (FE-Conv) to extract multi-scale frequency domain features and focuses on diagnostically critical areas through an attention mechanism. Additionally, to handle anatomical variations and structural changes, the model incorporates Deformable Convolution to enhance the representation of key features. DBSE-Net uses a dual-branch decoder to effectively decouple the segmentation and classification tasks. The Spatial decoder accurately segments structures such as the intervertebral disc, spinal canal, and spinous process, while the Semantic decoder refines the morphological features of the intervertebral disc for herniated and non-herniated disc classification. On the LA-MRI dataset, DBSE-Net achieves 91.95% mean Dice Similarity Coefficient (mDSC) and 85.47% mean Intersection-over-Union (mIoU) for segmentation, and 97.67% accuracy (ACC) and 98.67% area under the curve (AUC) for classification, outperforming state-of-the-art methods. These results demonstrate that DBSE-Net holds strong potential for intelligent diagnosis in LDH MRI and can be extended to 3D medical image analysis tasks.