【作者】 胡定一； 郑钰山； 姜志国； 谢凤英；

【Author】 HU Dingyi;ZHENG Yushan;JIANG Zhiguo;XIE Fengying;Image Processing Center,School of Astronautics,Beihang University;Beijing Advanced Innovation Center for Biomedical Engineering,Beihang University;

【机构】 北京航空航天大学宇航学院图像中心； 北京航空航天大学医工交叉创新研究院生物医学高精尖创新中心；

【摘要】 数字病理图像分析是当前癌症诊断中重要的辅助诊断方法,数字病理图像是指经过染色处理的病理切片在显微镜下成像后得到的数字图像。目前,数字病理图像诊断最常见的方法有:图像分割、图像分类和图像检索等。其中图像分割、分类可以提供直观的辅助诊断结果,但难以提供明确的诊断依据。相比之下,基于内容的图像检索可以在数据库中查找与目标患者相似的病理图像及其病例,为医生诊断提供更加丰富的辅助信息。现有数字病理图像检索方法通常借鉴自然场景图像检索模型,通过把WSI切分成图像块(Patch)的方式建立数据库,进而以图像为单位完成检索。该方法在实际应用中取得了一定的辅助诊断效果,然而病理图像是以全切片图像(Whole Slide Image,WSI)形式储存的,与自然图像有较大差异,导致现有方法难以满足医生的检索要求,其主要原因涉及以下两个方面:首先,癌变的区域通常以病变的区域表征病人的病情,而非固定大小的图像块,因此在检索时医生通常会根据诊断需要划定目标区域,并期望获得与该区域具有相似信息的历史病例,现存检索框架尚不能满足这一需求。其次,由于不同病人的切片大小存在较大差异,使切片包含的信息量差别可达1～2个数量级,致使返回的图像通常会忽略一些信息量小的切片、反复命中同一切片内相似的图像块、返回结果存在大量冗余信息、病例库利用率较低等一系列问题。本文针对上述问题,提出了一个针对区域进行检索的自上而下的深度哈希病理全切片图像检索方法,即首先从切片层面对检索结果进行排序,最后返回区域结果。本文主要从网络设计和检索策略两个方面进行优化,首先利用深度卷积神经网络对于大数据量的优秀的处理能力,搭建了以DenseNet为底层结构的神经网络,然后使用哈希函数对特征进行编码,以二进制码的形式进行储存,以实现大规模数据下的快速检索。网络采用哈希损失函数,并同时考虑了切片内与切片间相似度的差异,有效缓解了多次检索返回同一切片内容的情况。在检索策略方面,首先利用网络得出查询图像的组合二值编码,之后对于检索结果按照切片信息的丰富程度进行排序,在每一个切片中采用非极大值抑制的方法对相似的图像块进行融合最终获得目标区域结果。另一方面,设计了一种新的检索信息量评价指标,并对结果进行了评估。对本文的方法在ACDC_lungHP数据集上完成了验证,并与目前数字病理图像方面通用的检索方法进行了对比,通过实验验证了本方法可以挖掘更丰富的切片信息,在大规模数据集上实现高效、快速的检索,返回给医生更加灵活的区域图像。更多还原

【Abstract】 Digital histopathology image plays an important role in computer-aided cancer diagnosis.Digital histopathology image is obtained by imaging the stained slide under the microscope.Generally,there are three main categories in the domain of CAD:segmentation,classify and retrieval.The segmentation and classify cannot apply certain diagnostic basis.By contrast,the content-based image retrieval can retrieve the most similar images and the corresponding diagnosis information from the database.The retrieval can provide more information and guidance for the clinical diagnosis than the other two method.The current retrieval methods are mainly introduced from nature scene image retrieval,which has ignored the characteristic of histopathology image.To realize region retrieval from WSIs,the current retrieval method usually divides the WSI into individual patches and then recalls the stable size patches.This method has achieved promising retrieval performance.However,the histopathology image is usually stored in the form of a Whole Slide Image(WSI),which is very different from the normal natural image.It makes the current retrieval method cannot meet the requirement of pathologist.There are still two main issues:First of all,cancerous tissue generally appears as irregular regions according to the patient’s condition,rather than a fixed-size image patches,and the doctor usually enters a target region for retrieval rather than patches.Secondly,because the tissue size from different cases can be quite different,the amount of information contained in the training slides varies greatly.Some slides even suffer from two orders of magnitude difference.The slides with small amount of information are usually ignored in the retrieval.It causes that retrieved results repeatedly hits similar image patches in the same slide and returns a large number of redundant results.Meanwhile,the usage of the retrieval dataset is incomplete.Motivated by the aforementioned limitations,we proposed a noveltop-down slide-level deep hashbased retrieval method.Different from the current retrieval framework based on patch dataset,the returned regions in our method are sort based on the global similarity of the slides.We mainly make optimizes from two aspects:network design and retrieval strategy.Firstly,we built a network with DenseNet as the base structure for the powerful capacity of the deep convolutional neural network in huge data sets.Then,we apply the hash function to encoding the feature and store it in the form of binary code.The network is trained through hash loss and a designed slide-level loss to limit the case of repeatedly recall the same slide information.In terms of retrieval strategy,the network firstly obtains the combined binary coding of the query region image.Next,we sort the retrieval results according to the amount of the slide information.Then,we uses Non-maximum Suppression(NMS) in each slide to merge similar image patches,obtaining the fusion ultimately results as the returned region.Moreover,we design a new evaluation metric for the slide-level utilization to evaluate the results.The method of this paper was validated on the ACDC-lungHP dataset and compared with the current retrieval methods of digital histopathology images.The experimental results have demonstrated the proposed method is effective on large-scale histopathology image retrieval.更多还原