【作者】 徐黎明；

【导师】 吴健；

【作者基本信息】 浙江大学 ， 计算机技术（专业学位）， 2022， 硕士

【摘要】 肋骨骨折是临床上最为常见的胸部损伤。由于人体肋骨数量多且排列紧密,人工诊断肋骨骨折是一项劳动强度大、专业要求高的任务,因此计算机辅助肋骨骨折诊断显得尤为重要。肋骨骨折的辅助诊断算法需要完成两个连续的子任务:(1)肋骨骨折的检测,用于自动识别和定位肋骨骨折;(2)骨折病灶区域的分割,用于在检测的基础上进一步分割出骨折病灶的大小和空间形态,方便医生确定后续的治疗方案,提高医疗效率。目前,基于深度学习的肋骨骨折检测和分割方法主要存在两个不足:首先,肋骨骨折的空间形态非常复杂,这给深度学习模型的特征提取带来了挑战,形态复杂或是病变特征不明显的骨折容易被漏检和误判;其次,深度学习算法非常依赖标注数据的数量和质量,但肋骨骨折的标注成本高昂,不充分或者分布不够广泛的标注数据容易导致深度学习模型的泛化性能变差。为了实现精确且鲁棒的肋骨骨折检测和分割,本文提出了基于自监督学习与形态感知的肋骨骨折检测和分割方法,该方法使用级联的两个深度学习模型分别完成肋骨骨折的检测任务和骨折病灶区域的分割任务。针对肋骨骨折的检测任务,本文提出了基于像素级自监督对比学习的目标检测模型(Frac Det Net),该检测模型在大量无标注的肋骨数据上进行自监督预训练,有效解决了检测模型的泛化性问题。更重要的是,像素级的自监督对比学习增强了检测模型对人体肋骨形态的空间敏感性,这使得检测模型能进行更加有效的特征提取,有助于提升模型对复杂形态肋骨骨折的识别能力。针对骨折病灶区域的分割任务,本文提出了基于空间形态感知的多任务分割模型(Frac Seg Net),该分割模型在逐像素分割的基础上引入了边界预测的辅助任务,辅助任务用边界约束的形式对骨折病灶的复杂边界进行建模,实现了更加精准的骨折病灶分割。最后,公开数据集和私有数据集的实验结果证明了上述两个模型的有效性。对比不使用自监督预训练的基线模型,Frac Det Net将公开和私有数据集上的检测召回率分别提升了3.4%和7.6%,Frac Seg Net也分别取得了6.1%和4.7%的分割Dice指标提升。临床实验表明,本文方法在诊断的准确率和速度上均高于临床医师,并且深度学习算法与医师的交互式诊断能取得更好的检测和分割结果,这说明算法具有临床使用价值。更多还原

【Abstract】 Rib fractures are the most common clinical chest injuries.Manual identification of rib fractures is a labor-intensive and professional task due to the large amount and tight space arrangement of human ribs,which highlights the significance of computer-aided automatic rib fractures diagnosis.The computer-aided rib fracture diagnosis includes two consecutive subtasks:(1)rib fracture detection,which is used to identify and localize rib fractures;(2)fracture lesion segmentation,which is used to delineate fracture lesions with the size and spatial shape on the basis of detection result and makes it convenient for doctors to determine the follow-up treatment plan and improve medical efficiency.Nowadays,there are two main shortages in deep learning(DL)based rib fracture detection and segmentation method.First,the spatial shape of the rib fracture is variable and complex,which brings challenge to feature extraction of DL models.Fractures with inconspicuous lesions or complex shape would be easily missed and misjudged;Second,DL algorithms highly rely on the quantity and quality of the data annotations.Insufficient amount or unbalanced distribution of the dataset may lead to poor generalization performance of DL models as the high cost of the rib fractures labeling.To obtain accurate and robust rib fracture diagnosis results,we propose a novel rib fracture detection and segmentation method via self-supervised learning and shape-aware model,which contains two cascaded DL models for the rib fracture detection task and the fracture lesion segmentation task respectively.For the rib fracture detection task,we propose a pixel-level self-supervised contrastive rib fracture detection network(Frac Det Net),which utilizes self-supervised pre-training on massive unlabeled rib CT images to improve the generalization of the detection model.Furthermore,such contrastive learning strategy enhances the spatial sensitivity of human ribs of the model and enables more effective feature extraction,which is beneficial for the detection model to locate rib fractures with complex shapes.For the fracture lesion segmentation task,we propose a shape-aware multi-task rib fracture segmentation network(Frac Seg Net).In spite of the basic segmentation task,Frac Seg Net includes an auxiliary task of boundary prediction,which is trained under the form of boundary constraints for the complex fracture lesions and contributes to accurate boundary segmentation result of the target fracture lesions.Finally,experimental results on public and private datasets demonstrate the effectiveness of our method.Compared with the baseline methods without self-supervised pre-training,Frac Det Net improves the recall of detection on public and private datasets by 3.4%and 7.6%respectively.Frac Seg Net also achieves 6.1%and 4.7%improvement in the dice coefficient of the segmentation result.Clinical experiments show that our method outperforms clinicians in terms of the diagnostic accuracy and speed.The joint diagnosis of doctors and DL algorithm can further boost detection and segmentation results,which shows that our algorithm is of great clinical value.更多还原