Dual Attention Driven Lumbar Magnetic Resonance Image Feature Enhancement and Automatic Diagnosis of Herniation

TL;DR

The paper tackles automated lumbar disc herniation (LDH) diagnosis from MRI by developing a dual-attention framework that leverages multi-plane (T1/T2) data. It combines data augmentation and a ResNeXt-based backbone with CBAM and SEBlock to extract discriminative features, culminating in an MLP classifier that achieves an AUC of 0.969 and an ACC of 0.949. The approach aims to reduce reliance on radiologists and enable efficient LDH assessment in primary care, showing robustness across anatomical locations and improved generalization. Limitations include single-source data and lack of multimodal fusion, with future work targeting multi-center validation and integration of additional MRI modalities.

Abstract

Lumbar disc herniation (LDH) is a common musculoskeletal disease that requires magnetic resonance imaging (MRI) for effective clinical management. However, the interpretation of MRI images heavily relies on the expertise of radiologists, leading to delayed diagnosis and high costs for training physicians. Therefore, this paper proposes an innovative automated LDH classification framework. To address these key issues, the framework utilizes T1-weighted and T2-weighted MRI images from 205 people. The framework extracts clinically actionable LDH features and generates standardized diagnostic outputs by leveraging data augmentation and channel and spatial attention mechanisms. These outputs can help physicians make confident and time-effective care decisions when needed. The proposed framework achieves an area under the receiver operating characteristic curve (AUC-ROC) of 0.969 and an accuracy of 0.9486 for LDH detection. The experimental results demonstrate the performance of the proposed framework. Our framework only requires a small number of datasets for training to demonstrate high diagnostic accuracy. This is expected to be a solution to enhance the LDH detection capabilities of primary hospitals.

Figures (7)

• Figure 1: MRI of LDH. (a), The Sagittal Plane of MRI. (b), The axial os plane of MRI. The red box indicates the location of the LDH lesion.
• Figure 2: Different types of MRI. (a), T1-weighted images of the Sagittal Plane, reflecting differences in longitudinal relaxation times of tissues, can clearly show anatomical structures. (b), T2-weighted images of the Sagittal Plane, reflecting differences in tissue transverse relaxation times, can be used to visualize soft tissue lesions. (c), T2-weighted images of the Axial Plane.
• Figure 3: Overview of the structure. The architecture is structured into two primary components: data preprocessing and model training. The preprocessing stage applies data augmentation to MRI images, including rotation, inversion, panning, scaling, and noise injection, yielding 3,070 augmented training samples. Afterwards, the images of the same patient were divided into groups by setting an additional marker. In model training, data is processed through a CNN layer, followed by feature extraction via a dual-attention module integrating channel and spatial attention. Final LDH predictions are generated through a hierarchical network of Conv-block, Identity block, and a Multi-Layer Perceptron (MLP). The Conv-block and Identity block are repeated 4 times, and the Identity block appears 2 times, 3 times, 5 times, and 3 times each time.
• Figure 4: Channel attention block. It obtains the attention features of each channel through three parts: squeeze, excitation, and scale.
• Figure 5: Grouped convolutions used in Conv-block and Identity block. Input features are partitioned into distinct n groups for parallel convolution operations. Finally, the output features are combined.