A Robust Detection Network Utilizing Neural Memory Ordinary Differential Equations for Vertebral Fractures Diagnosis

Clinicians face challenges in diagnosing complex vertebral fractures, relying on a time-consuming manual classification process that hampers standardization. There is an urgent need for automation. Current automated methods encounter significant limitations, including overfitting due to insufficient data, the laborious and subjective nature of data annotation impeding label acquisition, and imbalances in sample numbers compromising classification accuracy across diverse fracture types. To address these challenges, we propose a one-stage, high-accuracy localization, and classification neural network, denoted as DEEP-Spine, empowered by the recent proposed nmODE (neural memory ordinary differential equations). DEEPSpine represents a cutting-edge solution designed to augment nonlinear capabilities, effectively addressing network overfitting and optimizing network training while ensuring computational efficiency. Additionally, we propose an innovative automatic labeling method to alleviate the manual burden on doctors and generate a comprehensive training dataset. Furthermore, we introduce a novel mechanism for constructing a sample set to mitigate the issue of imbalanced sample sizes and enhance classification accuracy. To validate the effectiveness of our approach, we collected CT images of 287 vertebral fracture cases from West China Hospital of Sichuan University and conducted experiments using Intersection over Union (IoU), Accuracy, Precision, Recall, and F1-score metrics on the constructed dataset. The experimental results demonstrate the validity and effectiveness of our model.