Optimizing spinal cord lesion segmentation using hierarchical classification and U-NET based segmentation model

Abstract

Rapid and accurate identification of spinal cord lesions is crucial in emergency medical care; however, current automated analysis systems face significant challenges in terms of processing speed and accuracy. This study presents a novel approach that combines hierarchical classification with segmentation models for automated spinal cord lesion analysis that can operate without human intervention, from image selection to lesion segmentation. We developed and evaluated a multistage classification system utilizing state-of-the-art efficient vision models, including EfficientVit, MobileNetV4, and RepVit, and compared it with a ResNet baseline. We incorporated a transfer learning strategy for lesion segmentation, leveraging pre-trained models to improve lesion boundary delineation. In addition, we introduced a synthetic lesion augmentation framework to address data scarcity and improve the generalization of the model. The system was trained and validated using six public datasets, including RadImageNet (1.4M images), and two private hospital datasets comprising 351 patients with spinal MRI scans. Our hierarchical classification approach achieved an end-to-end F1 score of 0.8357 using EfficientVit, whereas our optimized U-Net segmentation model attained a mean Dice score of 0.7527, surpassing previous approaches. Notably, the system achieved a ninefold increase in processing speed, reducing the lesion segmentation analysis time from 9.86 to 1.04 s per series and achieving a classification inference speed of 11.47 ms per image. Our findings suggest that this approach may support efficient automated spinal cord lesion analysis in resource-constrained clinical settings, potentially aiding diagnosis in emergency care, supporting radiologist workflow, and increasing the accessibility of spinal cord lesion analysis.