基于深度神经网络的骨质疏松自动检测方法

Automatic Detection Method for Osteoporosis Based on Deep Neural Networks

摘要: 本研究聚焦于自动检测老年人群中常见的骨质疏松性椎体骨折(OVF)。OVF在老年人中普遍存在，常导致严重的个人痛苦和社会经济负担。由于早期OVF可能无症状，导致其常被忽视，未能及时诊断和报告，从而影响患者的预后。本研究旨在利用深度卷积神经网络(CNN)从CT扫描片段中提取放射学特征，并通过特征聚合模块进行最终诊断，以提高OVF的早期检测率。在方法上，本研究探索长短期记忆网络(LSTM)对于提升系统诊断性能的表现。本研究在1432个CT扫描(包含10,546个2D矢状视图图像)上训练和评估了系统性能。结果表明，在129个CT扫描的保留测试集上，系统达到了89.2%的准确率和90.8%的F1分数。这一表现与实际临床环境中的执业放射科医师相当。本研究开发的自动检测系统通过预先筛查常规CT检查，并在放射科医生审阅前标记出可疑案例，有望协助和改善临床环境中的OVF诊断。这项技术的重要性在于，它不仅能提高诊断的准确率，还能减少放射科医生的工作负担，筛查出更多未被诊断的病例，优化医疗资源的分配。

Abstract: The study focuses on the automatic detection of Osteoporotic Vertebral Fractures (OVF), which are common in the elderly population. OVF is prevalent among the elderly and often leads to severe personal suffering and significant social-economic burden. Since early OVF may be asymptomatic, it is often overlooked and not diagnosed or reported in time, which affects patient prognosis. This study aims to utilize Deep Convolutional Neural Network (CNN) to extract radiological features from CT scan slices, and to use a feature aggregation module for final diagnosis to improve the early detection rate of OVF. In terms of methodology, this study explores the performance of Long Short-Term Memory (LSTM) networks in enhancing the diagnostic performance of the system. The system was trained and evaluated on 1432 CT scans (containing 10,546 2D sagittal view images). Results indicate that on the 129 CT scans from the held-out test set, the system achieved an accuracy of 89.2% and an F1 score of 90.8%. This performance is comparable to that of practicing radiologists in a clinical environment. The automatic detection system developed in this study can assist in pre-screening routine CT scans and flagging suspicious cases before radiologists’ review, potentially aiding and improving OVF diagnosis in clinical settings. The significance of this technology lies in its ability to not only improve diagnostic accuracy but also reduce radiologists’ workload, screen more undiagnosed cases, and optimize the allocation of medical resources.