HRNet-ACAP-Offset: a novel framework for localizing posterior edge landmark points in 3D intervertebral disc MRI

Locating the posterior edge points of intervertebral discs is essential for evaluating the spatial relationship between the discs and adjacent neural structures, which is of great clinical significance for diagnosing disc protrusion. However, manual annotation of these points on MRI scans is time-consuming and labor-intensive. To overcome the limitations of low efficiency and accuracy in 3D posterior edge point localization, this study proposes a novel HRNet-ACAP-Offset framework.

This framework adopts HRNet as the backbone and integrates a refined attention mechanism module (ACAP). Specifically, it employs a joint training paradigm with two task outputs and three loss functions. A 3D heatmap is utilized to regress the position of each landmark, while a landmark deviation estimation (Offset) method is adopted to correct the predicted positions. Furthermore, the regression of both positions and deviations is supervised by the original coordinate labels.

Experimental results on a clinical dataset demonstrate that the proposed method achieves an average localization error of 3.84 mm (2.25 pixels), outperforming standard baseline architectures (e.g., vanilla HRNet and ResNet34). Furthermore, ablation studies validate that the multi-loss joint training strategy and the offset estimation module significantly enhance the localization accuracy of the posterior edge points for this task.

The proposed HRNet-ACAP-Offset framework realizes efficient and high-precision automatic localization of intervertebral disc posterior edge points. It effectively solves the low efficiency and accuracy bottlenecks of traditional manual annotation and conventional model methods, possessing great potential for auxiliary clinical diagnosis of disc protrusion and related spinal diseases.