9. Detection of cervical ossification of the posterior longitudinal ligament with a Dual-Stage Attention and Multi-Scale Feature Fusion Network

Abstract

BACKGROUND CONTEXT

Cervical ossification of the posterior longitudinal ligament (C-OPLL) is a significant contributor to cervical myelopathy and presents surgical challenges due to its complex pathology. Magnetic resonance imaging (MRI) is the standard modality for evaluating neural compression in cervical spine disorders. However, in cases involving ossified lesions such as C-OPLL or partially calcified herniated intervertebral discs (HIVD), MRI often falls short in precisely identifying calcified structures. Computed tomography (CT) is required for definitive diagnosis but is associated with increased radiation exposure and higher costs.

PURPOSE

Recent advances in artificial intelligence (AI)-based imaging techniques offer a novel solution to enhance MRI’s diagnostic capability. By utilizing deep learning models to improve MRI images, AI-driven approaches could potentially reduce the reliance on CT scans, minimizing radiation exposure and associated costs while maintaining diagnostic accuracy. This study introduces a deep learning model designed to enhance MRI-based detection of C-OPLL and improve diagnostic workflows.

STUDY DESIGN/SETTING

We propose a deep learning framework, the Dual-Stage Attention and Multi-Scale Feature Fusion Network (DAFNet), for detecting C-OPLL using MRI. The model consists of two stages: segmentation and detection. In the segmentation stage, a dedicated segmentation model extracts the cervical spine region (C2–C6) to enhance image features. The refined images are then processed using the U-Coordinate Attention Mobile Inverted Bottleneck Convolution Network (U-CAMBNet), a model based on U-Net incorporating Agile Inverted Bottleneck Convolution (ABconv) and Coordinate Attention mechanisms. These enhancements enable improved feature extraction, allowing for precise classification and localization of ossified lesions.

PATIENT SAMPLE

N/A

OUTCOME MEASURES

N/A

METHODS

We propose a deep learning framework, DAFNet, for detecting C-OPLL using MRI. The model consists of two stages: segmentation and detection. In the segmentation stage, a dedicated segmentation model extracts the cervical spine region (C2–C6) to enhance image features. The refined images are then processed using the U-CAMBNet, a model based on U-Net incorporating ABconv and Coordinate Attention mechanisms. These enhancements enable improved feature extraction, allowing for precise classification and localization of ossified lesions.

RESULTS

Our AI-driven model demonstrated high diagnostic performance, achieving an accuracy of 0.99, precision of 0.76, recall of 0.77, and an F1 score of 0.76. These metrics surpass those of existing studies, highlighting the model’s potential to enhance MRI-based detection of C-OPLL and reduce dependence on CT imaging.

CONCLUSIONS

The application of AI-driven deep learning models in MRI analysis offers a promising alternative to traditional CT-based diagnostics for C-OPLL. By improving MRI-based detection accuracy, this approach can mitigate radiation exposure risks and alleviate financial burdens while ensuring diagnostic precision. AI-assisted imaging advancements could redefine cervical spine pathology management and optimize resource allocation in clinical practice.

FDA Device/Drug Status

This abstract does not discuss or include any applicable devices or drugs.