医学目标检测是一类重要的医学影像分析任务。随着深度学习方法在自然图像任务上取得巨大成功，相关方法也被迁移至医学目标检测任务中。由于医学数据集规模通常较小，且标注质量难以保证，通用目标检测模型的性能常常难以满足临床诊断的要求。与此同时，由于医学影像主要针对人体器官，普遍存在对应的先验知识，这些先验知识能够提供大量的信息和多维度的约束。所以，在神经网络中实现医学先验知识的有效嵌入是提升目标检测性能的重要途径。论文对医学目标检测场景中的先验知识进行系统性建模，并探索与神经网络融合的具体方法。医学场景中的待检测目标通常具有先验的形状大小和特定的存在区域，体现出目标自身的空间约束。目标的类别标签也常常含有丰富的信息，表现为类别标签之间的联系。论文基于目标的空间信息和标签信息，提出了级联学习结构，通过对原始任务分而治之的方式，降低了每一级子任务的处理难度。进一步，论文提出了多任务联合学习模型，同时建模多个子任务，依靠子任务之间的约束关系提高模型的稳健性，实现了性能的有效提升。大量医学影像中包含多个待检测目标，这些目标经常存在位置上的相互约束。论文基于目标之间的位置关系，提出了迭代优化方法，通过优化能量函数对模型输出的多目标检测结果进行校正。进一步，论文提出了目标关系模型，通过基于自注意力机制设计的目标关系单元，主动学习目标之间的位置模式。训练得到的神经网络模型能够对目标之间的位置关系有效建模。在临床诊疗过程中，为了全面获取人体器官的信息，针对同一个目标常会进行不同角度的成像。论文基于多视角医学影像的对应关系，提出了多视角融合模型。模型实现了多阶段的特征融合，针对多路输入图像，设计了图像级融合模块，共享多通道的骨干神经网络参数，并融合多视角的特征图。针对多路目标，设计了目标级融合模块，建模多视角中的目标对应信息。论文研究工作表明，在医学目标检测任务中，融合医学先验知识的深度学习模型，能够取得明显的性能提升。因此，论文提出的方法为先验知识在医学目标检测任务中的挖掘与建模提供了有效可行的技术路线。这些方法具有互补性，能够推广到很多类似任务中，并为医生提供可靠的临床诊断支持。
Medical object detection is a kind of important medical image analysis task. Deep learning methods have achieved great success in natural image tasks, and they are also used in medical object detection tasks. For the reason that the number of medical data is often insufficient and the quality of labeling cannot be assured, common object detection models usually cannot achieve the required performance of clinical diagnosis. Meanwhile, aiming to human organs, medical images always have relevant prior knowledge, which can provide much information and multi-dimension constraints. Therefore, inserting the medical prior knowledge into neural networks is an important way to improve the object detection performance. This paper models prior knowledge systematically in medical object detection scenes, and researches the relevant methods to combine medical prior knowledge with neural networks.The objects in medical scenes always have prior shape, size and existing areas, which indicates the space constraints of single object. Besides, different object labels also have relationship to each other, showing the correlation constraints between object labels. This paper researches the way to utilize these constraints, and proposes the cascade structure, which decreases the complexity of sub-tasks by using the divide-and-conquer method for original task. Furtherly, this paper proposes the multi-task learning model to handle all the sub-tasks simultaneously. This paper improves the model robustness by using the constraints between sub-tasks, and gets performance gain in a large margin.There are always multiple objects in medical images, and these objects can provide location constraints for each other. Based on the relationship between medical objects, this paper utilizes an energy function to refine the multi-objects output by neural networks, and improves the detection performance by iterative optimization. Furtherly, this paper proposes the object correlation model to learn the location pattern between objects, by designing the proposal correlation unit based on self-attention mechanism, so that the relationship between objects can be effectively modeled by neural networks.In diagnosis process, to get full information, the same objects are always imaged in different views. Based on the correspondence of multi-view medical images, this paper proposes the multi-view fusion model, which achieves multi-stage feature fusion. For multi-channel input images, this paper proposes the image-level feature fusion module to combine the information in feature maps of different views. For multi-channel detected objects, this paper designs the object-level feature fusion module, which can model the correspondence of objects in different views, and improve the detection performance furtherly.For medical object detection tasks, it shows that the deep learning models can achieve large performance improvement by combining with medical prior knowledge. Therefore, this paper proposes a feasible and efficient technique route for the utilization of prior knowledge in medical object detection tasks. The methods proposed in this paper have strong complementarity and can extend to many similar tasks, which can provide reliable assistance for clinicians.
