文献速递:深度学习肝脏肿瘤诊断---基于深度学习的肝细胞结节性病变在整片组织病理图像上的分类

本文主要是介绍文献速递:深度学习肝脏肿瘤诊断---基于深度学习的肝细胞结节性病变在整片组织病理图像上的分类,希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!

Title 

题目

Deep Learning-Based Classification of Hepatocellular Nodular Lesions on Whole-Slide Histopathologic Images

基于深度学习的肝细胞结节性病变在整片组织病理图像上的分类

Background 

背景

Hepatocellular nodular lesions (HNLs) constitute a heterogeneous group of disorders. Differential diagnosis among these lesions, especially high-grade dysplasticnodules (HGDNs) and well-differentiated hepatocellular carci noma (WD-HCC), can be challenging, let alone biopsy speci mens. We aimed to develop a deep learning system to solve these puzzles, improving the histopathologic diagnosis of HNLs (WD-HCC, HGDN, low-grade DN, focal nodular hyperplasia,hepatocellular adenoma), and background tissues (nodularcirrhosis, normal liver tissue).

肝细胞结节性病变(HNLs)构成了一个异质性疾病群。这些病变之间的鉴别诊断,特别是高级别发育不良结节(HGDNs)与良性分化的肝细胞癌(WD-HCC),可能具有挑战性,更不用说活检样本了。我们旨在开发一个深度学习系统来解决这些难题,以提高HNLs(WD-HCC、HGDN、低级别DN、局灶性结节性增生、肝细胞腺瘤)以及背景组织(结节性肝硬化、正常肝组织)的组织病理诊断。

Conclusions

结论

We first developed a deep learning diagnostic model for HNLs, which performed well and contributed to enhancing the diagnosis rate of early HCC and risk stratification of patients with HNLs. Furthermore, HnAIM had significant ad vantages in patch-level recognition, with important diagnostic implications for fragmentary or scarce biopsy specimens.

我们首次开发了一个用于HNLs的深度学习诊断模型,该模型表现良好,并有助于提高早期HCC的诊断率和HNLs患者的风险分级。此外,HnAIM在补丁层面识别方面具有显著优势,对于零碎或稀缺的活检样本具有重要的诊断意义。

Results

结果

We obtained 213,280 patches from 1115 whole-slide images of 738 patients. An optimal model was finally chosen based on F1 score and area under the curve value, named hepatocellular-nodular artificial intelligence model (HnAIM), with the overall 7-category area under the curve of 0.935 in the independent external validation cohort. For biopsy specimens, the agreement rate with sub specialists’ majority opinion was higher for HnAIM than 9 pa thologists on both patch level and whole-slide images level.

我们从738名患者的1115张整片幻灯片图像中获得了213,280个补丁。基于F1得分和曲线下面积值,最终选择了一个最优模型,命名为肝细胞结节性人工智能模型(HnAIM),在独立外部验证队列中,7类别的曲线下面积为0.935。对于活检样本,HnAIM与亚专家多数意见的一致率高于9名病理学家,无论是在补丁层面还是整片幻灯片图像层面。

Method

方法

The samples consisting of surgical and biopsy specimens were collected from 6 hospitals. Each specimen was reviewed by 2 to 3 subspecialists. Four deep neural networks (ResNet50, InceptionV3, Xception,and the Ensemble) were used. Their performances were eval uated by confusion matrix, receiver operating characteristic curve, classification map, and heat map. The predictive efficiency of the optimal model was further verified by comparing with that of 9 pathologists.

样本包括手术和活检标本,这些标本收集自6家医院。每个标本由2至3名亚专科医生审核。使用了四个深度神经网络(ResNet50、InceptionV3、Xception和集成网络)。它们的性能通过混淆矩阵、接收者操作特征曲线、分类图和热图进行评估。通过与9名病理医生的诊断结果进行比较,进一步验证了最优模型的预测效率。

Figure

图片

Figure 1. Data, study design, and HnAIM classification framework. Six independent data sets (Headquarters, Lingnan andYuedong Hospital of SYSUTH, SYSUFH, FSFPH, and GZFPH) were used in this study. (A) The Headquarters and YuedongHospital of SYSUTH data sets were used for developing a 7-category discriminative model, while the other 4 data sets wereused for the external testing. (B) The distribution of the samples for each type of liver nodule in model development (left) andindependent external validation (right). (C) Flow chart of the study. The data sets of the 7 categories were divided into thetraining (70%), validation (15%), and testing (15%) sets. Then, ROIs were labeled with green masks for each category. Patcheswere extracted from ROIs by OpenSlide library at  40 magnification with a size of 1024  1024. The training set was used totrain the ensemble model based on 3 basic models, while the validation set was used to fine-tune superparameters, such as learning rate, and the testing set used to evaluate models’ performances by confusion matrix, ROC curve, WSI-level classi-fication map, and patch-level heat map. Patches of liver biopsy specimens were predicted by the optimal model and areshown using a histogram, while the model’s referral decisions were compared with the ones made by different levels ofpathologists.

图1. 数据、研究设计和HnAIM分类框架。本研究使用了六个独立数据集(总部、岭南及SYSUTH的粤东医院、SYSUFH、FSFPH和GZFPH)。(A) 总部和SYSUTH的粤东医院数据集用于开发7类鉴别模型,而其他四个数据集用于外部测试。(B) 模型开发中(左)和独立外部验证中(右)各类型肝结节样本的分布。(C) 研究流程图。7类数据集被划分为训练集(70%)、验证集(15%)和测试集(15%)。然后,每个类别的感兴趣区域(ROIs)用绿色遮罩标记。通过OpenSlide库以40倍放大从ROIs提取1024×1024大小的补丁。训练集用于基于三个基础模型训练集成模型,验证集用于调整超参数,如学习率,测试集用于通过混淆矩阵、ROC曲线、WSI级分类图和补丁级热图评估模型性能。肝活检标本的补丁由最优模型预测,并通过直方图显示,而模型的转诊决定与不同级别的病理医生所做的决定进行比较。

图片

Figure 2. Performance of deep learning models. (A) Classification results are shown by confusion matrices on the internal testing set for Resnet50, Inception V3, Xception, and the Ensemble model. Numbers represent the number of patches classified correctly (diagonal) and incorrectly (off the diagonal). (B) The ROC curve and the AUC value on the internal testing set for models of Resnet50 (black line), Inception V3 (blue line), Xception (green line), and Ensemble (red line). The Xception and the Ensemble models both performed the best, with AUC values of 0.9991, indicating models were trained with high accuracy. (C) The ROC curve and AUC value on the independent external validation using the Ensemble model (HnAIM) in FSFPH, SYSUFH, GZFPH, and the entire external data set.

图2. 深度学习模型的性能。(A) 在内部测试集上,Resnet50、Inception V3、Xception和集成模型的分类结果通过混淆矩阵显示。数字代表正确分类(对角线上)和错误分类(对角线外)的补丁数量。(B) 在内部测试集上,Resnet50(黑线)、Inception V3(蓝线)、Xception(绿线)和集成模型(红线)的ROC曲线和AUC值。Xception和集成模型的表现最佳,AUC值为0.9991,表明模型具有高精度的训练。(C) 使用集成模型(HnAIM)在FSFPH、SYSUFH、GZFPH和整个外部数据集上的独立外部验证的ROC曲线和AUC值。

图片

Figure 3. WSI-level panoramicclassification map of surgicalsample: (A) WD-HCC, (B)HGDN, (C), LDN, (D), FNH, and(E) HCA. (Left) Original WSIs(original magnification  0.4).(Middle) Classification mapswere constructed frommodel’s predictions of corresponding patches. Colorsfrom blue to red meantdifferent liver lesions. For NC,LGDN, HGDN, and WDHCC,gradually deepening coloreven indicated increased degree of malignancy (labels: 2,5–7). The diagnostic labelswere as follows: 0 for background, 1 for NNL, 2 for NC, 3for HCA, 4 for FNH, 5 forLGDN, 6 for HGDN, and 7 forWDHCC. (Right) Pie charts

quantitatively show the percentage of different categoriesin each WSI.

图3. 外科样本的WSI级全景分类图:(A) WD-HCC,(B) HGDN,(C) LDN,(D) FNH,和 (E) HCA。(左) 原始WSIs(原始放大倍数0.4)。(中) 分类图根据模型对应补丁的预测构建。颜色从蓝色到红色表示不同的肝脏病变。对于NC、LGDN、HGDN和WDHCC,颜色的逐渐加深甚至表示恶性程度的增加(标签:2,5-7)。诊断标签如下:0代表背景,1代表NNL,2代表NC,3代表HCA,4代表FNH,5代表LGDN,6代表HGDN,7代表WDHCC。(右) 饼图定量显示每个WSI中不同类别的百分比。

图片

Figure 4. Performance of HnAIM in biopsy specimens and comparison with pathologists. (A) Patch-level histogram of biopsy specimens shows the model’s predictions for 7 categories, with a focus on cell morphologic features. The category with the largest proportion was regarded as the final classification. Agreement rates with the majority opinion of subspecialists for the HnAIM and pathologists (3 each for junior, intermediate, and senior pathologist) on 7 categories across (B) all 961 patches and (C) 30 WSIs of biopsy specimens. To represent the average level of each group, the agreement rate was shown as the mean value across 3 pathologists. The error bars represent the 95% CIs. Potential reasons for disagreements among pathologists with HnAIM may include inherent uncertainty in the 2-dimensional interpretation of a 3-dimensional specimen, ambiguity in diagnostic guidelines, the limited number of tissue samples, and cognitive factors such as anchoring.

图4. HnAIM在活检标本中的表现及与病理医生的比较。(A) 活检标本的补丁级直方图显示了模型对7个类别的预测,重点关注细胞形态特征。占比最大的类别被视为最终分类。HnAIM与亚专家多数意见的一致率以及(B)所有961个补丁和(C)30个活检样本WSI中7个类别的病理医生(初级、中级和高级各3名)的一致率。为代表每组的平均水平,一致率以3名病理医生的平均值显示。误差条表示95%置信区间。病理医生与HnAIM之间意见不一的潜在原因可能包括对三维标本二维解读的固有不确定性、诊断指南的模糊性、组织样本数量有限以及认知因素如锚定效应。

Table

图片

Table 1.Seven-Category Agreement With Subspecialists’ Majority Opinion of 9 Pathologists and Hepatocellular-NodularArtificial Intelligence Model Based on Patches and Whole-Slide Images of 30 Liver Biopsy Specimens

表1. 基于30个肝活检标本的补丁和整片图像的九名病理学家和肝细胞结节性人工智能模型与亚专家多数意见的七类别一致性

图片

Table 2.Lesion Characteristics of Patients With Indefinite Diagnoses after 3 Independent Reviews

表2. 经过三次独立审查后,诊断不确定的患者的病变特征

这篇关于文献速递:深度学习肝脏肿瘤诊断---基于深度学习的肝细胞结节性病变在整片组织病理图像上的分类的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!



http://www.chinasem.cn/article/908812

相关文章

SpringCloud动态配置注解@RefreshScope与@Component的深度解析

《SpringCloud动态配置注解@RefreshScope与@Component的深度解析》在现代微服务架构中,动态配置管理是一个关键需求,本文将为大家介绍SpringCloud中相关的注解@Re... 目录引言1. @RefreshScope 的作用与原理1.1 什么是 @RefreshScope1.

Python 中的异步与同步深度解析(实践记录)

《Python中的异步与同步深度解析(实践记录)》在Python编程世界里,异步和同步的概念是理解程序执行流程和性能优化的关键,这篇文章将带你深入了解它们的差异,以及阻塞和非阻塞的特性,同时通过实际... 目录python中的异步与同步:深度解析与实践异步与同步的定义异步同步阻塞与非阻塞的概念阻塞非阻塞同步

Redis中高并发读写性能的深度解析与优化

《Redis中高并发读写性能的深度解析与优化》Redis作为一款高性能的内存数据库,广泛应用于缓存、消息队列、实时统计等场景,本文将深入探讨Redis的读写并发能力,感兴趣的小伙伴可以了解下... 目录引言一、Redis 并发能力概述1.1 Redis 的读写性能1.2 影响 Redis 并发能力的因素二、

最新Spring Security实战教程之表单登录定制到处理逻辑的深度改造(最新推荐)

《最新SpringSecurity实战教程之表单登录定制到处理逻辑的深度改造(最新推荐)》本章节介绍了如何通过SpringSecurity实现从配置自定义登录页面、表单登录处理逻辑的配置,并简单模拟... 目录前言改造准备开始登录页改造自定义用户名密码登陆成功失败跳转问题自定义登出前后端分离适配方案结语前言

Java进阶学习之如何开启远程调式

《Java进阶学习之如何开启远程调式》Java开发中的远程调试是一项至关重要的技能,特别是在处理生产环境的问题或者协作开发时,:本文主要介绍Java进阶学习之如何开启远程调式的相关资料,需要的朋友... 目录概述Java远程调试的开启与底层原理开启Java远程调试底层原理JVM参数总结&nbsMbKKXJx

Redis 内存淘汰策略深度解析(最新推荐)

《Redis内存淘汰策略深度解析(最新推荐)》本文详细探讨了Redis的内存淘汰策略、实现原理、适用场景及最佳实践,介绍了八种内存淘汰策略,包括noeviction、LRU、LFU、TTL、Rand... 目录一、 内存淘汰策略概述二、内存淘汰策略详解2.1 ​noeviction(不淘汰)​2.2 ​LR

使用Python开发一个图像标注与OCR识别工具

《使用Python开发一个图像标注与OCR识别工具》:本文主要介绍一个使用Python开发的工具,允许用户在图像上进行矩形标注,使用OCR对标注区域进行文本识别,并将结果保存为Excel文件,感兴... 目录项目简介1. 图像加载与显示2. 矩形标注3. OCR识别4. 标注的保存与加载5. 裁剪与重置图像

Python与DeepSeek的深度融合实战

《Python与DeepSeek的深度融合实战》Python作为最受欢迎的编程语言之一,以其简洁易读的语法、丰富的库和广泛的应用场景,成为了无数开发者的首选,而DeepSeek,作为人工智能领域的新星... 目录一、python与DeepSeek的结合优势二、模型训练1. 数据准备2. 模型架构与参数设置3

Java深度学习库DJL实现Python的NumPy方式

《Java深度学习库DJL实现Python的NumPy方式》本文介绍了DJL库的背景和基本功能,包括NDArray的创建、数学运算、数据获取和设置等,同时,还展示了如何使用NDArray进行数据预处理... 目录1 NDArray 的背景介绍1.1 架构2 JavaDJL使用2.1 安装DJL2.2 基本操

最长公共子序列问题的深度分析与Java实现方式

《最长公共子序列问题的深度分析与Java实现方式》本文详细介绍了最长公共子序列(LCS)问题,包括其概念、暴力解法、动态规划解法,并提供了Java代码实现,暴力解法虽然简单,但在大数据处理中效率较低,... 目录最长公共子序列问题概述问题理解与示例分析暴力解法思路与示例代码动态规划解法DP 表的构建与意义动