作为脊柱图像分析的一种基础任务，脊柱关键点检测能够定位识别脊柱结构，是脊柱侧弯和椎体骨折等脊柱疾病的影像诊断和后续治疗的重要先决步骤。虽然基于深度学习的关键点自动检测算法能高效地提供较为准确的识别结果，节省手动脊柱定位的时间，但对于大多数现有方法，均存在一定数量违背解剖学先验的样本位置预测，即相邻椎骨关键点重叠或者间隔距离过大，导致算法在实际应用中的可解释性和可靠性下降。针对上述问题，本文系统性地研究解剖结构先验在基于概率热力图估计的脊柱关键点检测框架中的应用以提升算法识别关键点的定位精度和解剖学合理性。在显式定义两种有关椎骨间隔距离和空间排列顺序的解剖学先验后，本文将先验知识分别融入算法的训练损失、推理过程和评价指标中。首先，本文设计了与解剖结构先验相关的损失函数，其作为惩罚项加入训练目标函数，约束各椎骨在空间上均匀分布，起到正则化作用；其次，提出了一种有效的基于解剖学先验的推理方式，使用串行的方式逐次预测各椎体位置，以代替常用的并行提取方式；最后，设计了有关解剖学先验的量化评价指标来评估关键点检测结果的解剖结构合理性，作为关键点定位误差基本指标的有效补充以更全面地衡量算法表现。基于对关键点检测基本任务的研究，本文针对脊柱侧弯和骨质疏松性椎体压缩性骨折两种常见疾病实现两个智能化辅助诊断方案。首先，本文将基于解剖学先验的关键点检测算法应用在多医疗中心收集的正位X光片上用于脊柱侧弯辅助诊断。与主流相关方法相比，算法取得最优的关键点定位精度和更佳的脊柱弯曲Cobb角估计结果，并且在初步临床试验中展现出与专家医生接近的诊断精度。同时本文通过消融实验证明了设计的先验相关的训练惩罚项和顺序推理机制在降低定位误差和提升解剖结构合理性方面的作用，从而能提供更可靠的下游诊断应用。其次，本文还将关键点定位算法迁移到脊柱侧位CT扫描上作为检测发生骨折的椎骨的前提步骤，并结合CNN分类器搭建全自动椎体压缩骨折检测框架。通过在79个CT扫描序列上实验，展示出较优的定位识别和骨折分类性能。同时，在各对比方法中本算法同样取得了最低的椎骨定位误差，相应地在下游的骨折分类任务中取得更优的 ROC-AUC值。
As one of fundamental ways to interpret spine images, detection of vertebral landmarks is an informative prerequisite for further diagnosis and management of spine disorders such as scoliosis and fractures. Most existing deep learning-based methods for automatic vertebral landmark detection suffer from overlapping landmarks or abnormally long distances between % erroneous order of nearby landmarks against anatomical priors, and thus lack sufficient reliability and interpretability. To tackle the problem, this paper systematically utilizes anatomical prior knowledge in vertebral landmark detection. The thesis explicitly formulates anatomical priors of the spine, related to distances among vertebrae and spatial order within the spine, and integrates these geometrical constraints within training loss, inference procedure, and evaluation metrics. First, the thesis introduces an anatomy-constraint loss to regularize the training process with the aforementioned contextual priors explicitly. Second, the thesis proposes a simple-yet-effective anatomy-aided inference procedure by employing sequential prediction rather than a parallel counterpart. Third, the thesis provides novel anatomy-related metrics to quantitatively evaluate to which extent landmark predictions follow the anatomical priors, as is not reflected within the widely-used landmark localization error metric. Based on the anatomy prior knowledge-based vertebral landmark detection algorithm, this thesis builds two fully automatic frameworks to assist the clinical diagnosis of scoliosis and detection of osteoporotic vertebral fracture. (1) The thesis employs the localization framework on 1410 anterior-posterior radiographic images for computer-aided scoliosis diagnosis. Compared with two competitive baseline models, the thesis achieves superior landmark localization accuracy and comparable Cobb angle estimation for scoliosis assessment. Through an initial trial in the clinical routine and comparison with manual measurement by clinicians, our algorithm exhibits the potential for clinical usage. Besides, ablation studies demonstrate the effectiveness of designed components on the decrease of localization error and improvement of anatomical plausibility, thus benefitting the interpretability and reliability of downstream clinical applications. (2) The thesis exhibits effective generalization performance by transferring our detection method onto sagittal 2-D slices of spine CT scans, and with CNN-based classifier, this thesis realizes a fully-automated pipeline for vertebral-level fracture detection. The system has demonstrated relatively low localization errors and high ROC-AUC values on 79 scans. Moreover, attributed to the improvement of landmark detection, the performance of downstream compression fracture classification would be boosted compared to the other two landmark-based automatic fracture analysis systems.
