Enhancing LDD diagnosis with YOLOv9-AID: simultaneous detection of pfirrmann grading, disc herniation, HIZ, and Schmorl's nodules.

Abstract

This study develops an intelligent diagnostic model for LDD based on a novel YOLOv9-AID detection network and evaluates the impact of multiple innovative strategies on detection performance. A total of 222 adult patients who underwent lumbar MRI for low back pain or radicular leg pain were enrolled, yielding 1110 de-identified sagittal T2-weighted images (five per case). After excluding cases with prior spinal trauma, tumors, postoperative changes, congenital malformations, or severe artefacts, 202 cases (1,010 images) were randomly divided into training, validation, and internal test sets (8:1:1), while 20 cases (100 images) formed an external dataset for generalization assessment. The YOLOv9-AID model introduces three key enhancements to the baseline YOLOv9: a SlideLoss function to rebalance training weights between high- and low-quality samples; spatial-channel collaborative attention modules (SCSA) embedded at layers 5 and 11 to strengthen lesion feature extraction; and an ExtraDW-inspired redesign of the ResNCSPELAN4 module to boost precision and reduce parameter count. In the internal test set, the model achieved an mAP50 of 82.8% and an overall detection precision of 80.3%, with Schmorl's node detection at 92.9%, Pfirrmann grading accuracy at 93.3%, and disc herniation accuracy at 73.2% (an 8.4% improvement). Recall rates increased by approximately 5% on average, with Schmorl's node recall up 15.1%, Pfirrmann recall up 1.8%, and herniation recall improvements of up to 12.3%. External validation confirmed robust generalization, and detection rates for small lesions such as high-intensity zones and Schmorl's nodes significantly outperformed conventional methods. These results demonstrate that the YOLOv9-AID network, through the integration of SlideLoss and spatial-channel attention mechanisms, substantially enhances the accuracy and robustness of LDD detection on MRI and offers a promising tool to support clinical diagnosis efficiency and consistency.