基于X光片影像智能识别脊柱侧弯椎体边界的深度学习模型及方法研究

脊柱侧弯是最常见的脊柱畸形，青少年中最为高发，在10～16岁人群中发病率达3～4%。脊柱全长X光片上的影像学参数是脊柱侧弯诊断、评估、以及手术方案设计过程中的主要参考指标。这些指标的测量目前需要医生手动完成，而人工操作伴随着不可避免的人力资源消耗以及主观误差因素。以往研究中尝试用传统机器学习和模式识别算法进行椎体的自动化检测，但由于X光片成像条件限制（对比度较低、图像较模糊），以及椎体形态变异、肋骨遮挡等原因，算法识别效果不尽人意。本项目拟开发基于深度学习方法的网络模型，根据脊柱侧弯X光片影像数据“小样本、高噪声、弱标签”的特点，针对性地采用将数据增强、模型精调、引入专家知识库等方法相结合的综合性解决方案，提高网络模型的准确率和稳定性，实现X光片中脊柱侧弯椎体边界的自动化识别和检测，为进一步的影像学参数自动化测量奠定基础，最终实现脊柱侧弯的计算机辅助诊断评估及更精准的治疗。

Scoliosis represents the most common type of spinal deformity with prevalence as high as 3～4% found in the adolescent population of age 10～16. Imaging characteristics, e.g. the Cobb angle, in long-cassette spinal X-rays are instrumental in the diagnosis, evaluation, and surgical planning of patients with scoliosis. However, manual measurement of these parameters can be both time-consuming and unreliable in clinical practice. As shown by results from previous attempts using conventional machine learning and pattern recognition techniques, automated vertebrae detection was a challenging task due to low tissue contrast intrinsic to X-ray images, occlusion of the vertebral body by ribs, etc. In this study, we aim to propose a deep learning network architecture that can provide a fully automated framework for vertebral boundary recognition in long-cassette spinal X-rays. A combination of different techniques, including data amplification, model fine-tuning, etc., are utilized to improve on the state-of-the-art network model in order to mitigate the constraints of limited dataset size, weak label and high image noise. The optimized model may then facilitate accurate and robust vertebral boundary recognition and benefit the clinical workflow by providing a computer-aided decision-making tool for precise treatment-planning.