毕业生知识库

    病理检查是疾病特别是癌症诊断的金标准，与癌症的分级和治疗密切相关。传统的病理诊断过程是病理学家在显微镜下观察和研究组织切片以给出诊断结论。近年来，随着数字病理学诞生，病理切片能够保存为全切片数字病理图像，从而通过计算机辅助诊断方法进行诊断。过去几年来，大量的病理辅助诊断方法被提出并取得了良好的效果。然而，全切片病理图像同传统图像差异较大，其通常以多尺度的空间金字塔结构存储，包含在不同倍镜下扫描的图像，最大尺寸可达10万x10万像素。其高分辨率的特性致使主流的病理辅助诊断方法因需要应用切分采样的策略执行密集扫描而具有较大的计算量和运行负载，难以在实际临床中进行应用。当前迫切需要高效的解决方案，但是现有的研究在加速方面的工作较少且没有均衡优化资源消耗。因此，本文以实现精准且高效的病理诊断为目标，提出了一个新颖、快速且低资源消耗的病理诊断框架。

    本文的主要贡献归纳如下：

    1. 提出了一个略缩图分割与轻量化分类相结合的快速诊断模型。病理学家的检查模式和当前计算机辅助诊断的检查模式不同。前者聚焦于疑似肿瘤的感兴趣区域，后者则在应用切分采样的策略密集分析图像上的所有前景区域后给出诊断结论，整个过程非常耗时。本文模拟病理学家的检查模式，提出了一个全新的解决方案。该方案首先以高灵敏度的方式实现疑似肿瘤的感兴趣区域的分割，之后仅在这些感兴趣区域上应用高效的轻量化卷积神经网络分类模型以降低计算量和资源消耗，最终在速度、精度以及资源消耗三方面实现了综合优化。

    2. 提出了一个基于全卷积分类网络架构的快速诊断模型。上述提出的轻量化模型显著提高了推理速度、减少了资源消耗，但是其仍然具有在切分采样框架下由于应用卷积神经网络模型进行推理带来的重叠区域的冗余计算以及输入图像大小受限的问题。针对上述问题，本文针对整体解决方案进行统筹优化，基于提出的轻量化卷积神经网络分类模型进一步改进，设计了基于全卷积分类网络架构的快速诊断模型。通过充分利用全卷积模型进行高效的并行推理，本文提出的快速诊断模型突破了全切片病理图像诊断过程中的主要速度瓶颈，最大化地挖掘了速度提升的潜力，且在精度方面相比于改进前的卷积神经网络模型没有显著降低。

    3. 设计并构建了病理图像高效诊断原型系统。病理图像智能诊断能够突破传统病理诊断具有的耗时、耗力且一致性低的瓶颈。在上述研究成果的基础上，本文还构建了临床可用的智能诊断系统以快速给出肿瘤病灶区域的机器辅助诊断结果。针对不同的任务场景需求，能够辅助病理学家实现精准且快速的诊断。

    Pathological examination is the gold standard for the diagnosis of diseases, especially cancer, and is closely related to the grading and treatment of cancer. The traditional pathological diagnosis process is that pathologists observe and study tissue slides under a microscope to make a diagnosis. In recent years, with the birth of digital pathology, pathological slides can be saved as whole slide pathological images in digital form, so as to be diagnosed by computer-aided diagnosis. In the past few years, a large number of pathologically assisted diagnosis methods have been proposed and achieved good results. However, compared with traditional images, whole slide pathological images are quite different. They are generally stored in a multi-scale spatial pyramid structure containing images scanned under different magnification, with a maximum size of 100,000x100,000 pixels. Due to its high resolution, the existing mainstream pathological diagnosis method has a large amount of computation and operation load due to the application of patch sampling strategy to perform intensive scanning, which is difficult to be applied in actual clinical practice. There is an urgent need for efficient solutions, but existing studies do little to accelerate and do not optimise resource consumption evenly. Therefore, in order to achieve accurate and efficient pathological diagnoses, a novel, rapid and low resource consumption pathological diagnosis framework is proposed in this paper. 

    The main contributions of this paper are summarized as follows:

    1. A rapid diagnosis model combining region of interest (ROI) segmentation and lightweight classification is proposed. The pathologist's examination mode differs from that of current computer-aided diagnosis. The former focuses on the ROI of the suspected tumor, while the latter uses a strategy of patch sampling to intensively analyze all the foreground areas of an image to make a diagnosis. The whole process is very time-consuming. In this paper, a novel solution is proposed to simulate the pathologist's examination mode, which first realizes the segmentation of the ROI of suspected tumors in a highly sensitive manner, and then only applies the efficient lightweight convolutional neural network classification model to the ROI to reduce the computation and resource consumption. Finally, the comprehensive optimization is achieved in speed, accuracy and resource consumption.

    2. A rapid diagnosis model based on fully convolutional classification network architecture is proposed. The proposed lightweight model significantly improves the inference speed and reduces the resource consumption, but it still has the problems of redundant calculation of overlapping areas and limited input image size due to the application of convolutional neural network model for inference under the patch sampling framework. In view of the above problems, this paper made overall optimization for the overall solution. Based on the proposed lightweight convolutional neural network classification model, a rapid diagnosis model based on the fully convolutional classification network architecture is designed. By making full use of the fully convolutional model for efficient parallel inference, the proposed rapid diagnosis model overcomes the main speed bottleneck in whole slide pathological image diagnosis and maximizes the potential of speed improvement, without significantly reducing the accuracy compared with the previous convolutional neural network model.

    3. The prototype system for efficient diagnosis of pathological images was designed and constructed. The intelligent diagnosis of pathological images can break through the bottleneck of time-consuming, labor-consuming and low consistency of traditional pathological diagnosis. On the basis of the above research results, this paper also constructed a clinical intelligent diagnosis system to quickly give the machine-aided diagnosis results of the tumor lesion area. According to the requirements of different task scenarios, it can assist pathologists to achieve accurate and rapid diagnoses.