Pub Date : 2025-08-06DOI: 10.1016/j.jpi.2025.100465
Tony Yeung , Yi Zhang , Qianghua Zhou , Richard Burack
Evaluation of tumor infiltrating lymphocytes as recommended by current guidelines is largely based on stromal regions within the tumor. In the context of epithelial malignancies, the epithelial region and the epithelial–stromal interface are not assessed, because of technical difficulties in manually discerning lymphocytes when admixed with epithelial tumor cells. The inability to quantify immune cells in epithelial-associated areas may negatively impact evaluation of patient response to immune checkpoint therapies. Innovative spatial analysis techniques have emerged that can directly address challenges associated with quantification of lymphocytes in specialized regions like the interface. In this study, we apply supervised neighborhood clustering analysis (via an open-source application CytoMAP) to assess the spatial distribution of CD8+ T cells, CD8+ TIM3+ (T cell immunoglobulin and mucin-domain containing-3) exhausted T cells, and TIM3+ CD8- macrophages on a gynecological tumor microarray. Neighborhood clustering analysis is adept at objectively mapping the epithelial–stromal interface alongside the epithelial and stromal region of each tumor under a three-compartment model. When tumors are partitioned by the conventional two-compartment model (epithelial and stromal region only), the highest density of total CD8+ T cells is found in the stromal region in a slight majority of tumors. In contrast, the interface region surpasses both the epithelial and stromal region in holding the highest density of CD8+ T cells when this unique region is incorporated into the three-compartment model. Further subset analysis shows higher proportion of CD8+ TIM3+ exhausted T cells within the interface and epithelial region, as compared to CD8+ TIM3- T cells which span from the stroma to the interface. These results highlight the utility of implementing quantitative spatial technique and immune subset analysis in the assessment of tumor infiltrating lymphocytes, and underscore the potential significance of the under-reported tumor epithelial–stromal interface.
{"title":"Neighborhood clustering analysis to define epithelial–stromal interface for tumor infiltrating lymphocyte evaluation","authors":"Tony Yeung , Yi Zhang , Qianghua Zhou , Richard Burack","doi":"10.1016/j.jpi.2025.100465","DOIUrl":"10.1016/j.jpi.2025.100465","url":null,"abstract":"<div><div>Evaluation of tumor infiltrating lymphocytes as recommended by current guidelines is largely based on stromal regions within the tumor. In the context of epithelial malignancies, the epithelial region and the epithelial–stromal interface are not assessed, because of technical difficulties in manually discerning lymphocytes when admixed with epithelial tumor cells. The inability to quantify immune cells in epithelial-associated areas may negatively impact evaluation of patient response to immune checkpoint therapies. Innovative spatial analysis techniques have emerged that can directly address challenges associated with quantification of lymphocytes in specialized regions like the interface. In this study, we apply supervised neighborhood clustering analysis (via an open-source application CytoMAP) to assess the spatial distribution of CD8+ T cells, CD8+ TIM3+ (T cell immunoglobulin and mucin-domain containing-3) exhausted T cells, and TIM3+ CD8- macrophages on a gynecological tumor microarray. Neighborhood clustering analysis is adept at objectively mapping the epithelial–stromal interface alongside the epithelial and stromal region of each tumor under a three-compartment model. When tumors are partitioned by the conventional two-compartment model (epithelial and stromal region only), the highest density of total CD8+ T cells is found in the stromal region in a slight majority of tumors. In contrast, the interface region surpasses both the epithelial and stromal region in holding the highest density of CD8+ T cells when this unique region is incorporated into the three-compartment model. Further subset analysis shows higher proportion of CD8+ TIM3+ exhausted T cells within the interface and epithelial region, as compared to CD8+ TIM3- T cells which span from the stroma to the interface. These results highlight the utility of implementing quantitative spatial technique and immune subset analysis in the assessment of tumor infiltrating lymphocytes, and underscore the potential significance of the under-reported tumor epithelial–stromal interface.</div></div>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"19 ","pages":"Article 100465"},"PeriodicalIF":0.0,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144912219","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-08-01DOI: 10.1016/j.jpi.2025.100450
Peter Gershkovich
<div><div>Slides are data. Once digitized, they function like any enterprise asset: accessible anywhere, ready for AI, and integrated into cloud workflows. But in pathology, they enter a realm of clinical complexity—demanding systems that handle nuance, integrate diverse data streams, scale effectively, enable computational exploration, and enforce rigorous security.</div><div>Although the Digital Imaging and Communications in Medicine (DICOM) standard revolutionized radiology, it is imperative to explore its adequacy in addressing modern digital pathology's orchestration needs. Designed more than 30 years ago, DICOM reflects assumptions and architectural choices that predate modular software, cloud computing, and AI-driven workflows.</div><div>This article shows that by embedding metadata, annotations, and communication protocols into a unified container, DICOM limits interoperability and exposes architectural vulnerabilities. The article begins by examining these innate design risks, then challenges DICOM's interoperability claims, and ultimately presents a modular, standards-aligned alternative.</div><div>The article argues that separating image data from orchestration logic improves scalability, security, and performance. Standards such as HL7 FHIR (Health Level Seven Fast Healthcare Interoperability Resources) and modern databases manage clinical metadata; formats like Scalable Vector Graphics handle annotations; and fast, cloud-native file transfer protocols, and microservices support tile-level image access. This separation of concerns allows each component to evolve independently, optimizes performance across the system, and better adapts to emerging AI-driven workflows—capabilities that are inherently constrained in monolithic architectures where these elements are tightly coupled.</div><div>It further shows that security requirements should not be embedded within the DICOM standard itself. Instead, security must be addressed through a layered, format-independent framework that spans systems, networks, applications, and data governance. Security is not a discrete feature but an overarching discipline—defined by its own evolving set of standards and best practices. Overlays such as those outlined in the National Institute of Standards and Technology SP 800-53 support modern Transport Layer Security, single sign-on, cryptographic hashing, and other controls that protect data streams without imposing architectural constraints or restricting technological choices.</div><div>Pathology stands at a rare inflection point. Unlike radiology, where DICOM is deeply entrenched, pathology workflows still operate in polyglot environments—leveraging proprietary formats, hybrid standards, and emerging cloud-native tools. This diversity, often seen as a limitation, offers a clean slate: an opportunity to architect a modern, modular infrastructure free from legacy constraints. While a full departure from DICOM is unnecessary, pathology is uniquely position
{"title":"Wearing a fur coat in the summertime: Should digital pathology redefine medical imaging?","authors":"Peter Gershkovich","doi":"10.1016/j.jpi.2025.100450","DOIUrl":"10.1016/j.jpi.2025.100450","url":null,"abstract":"<div><div>Slides are data. Once digitized, they function like any enterprise asset: accessible anywhere, ready for AI, and integrated into cloud workflows. But in pathology, they enter a realm of clinical complexity—demanding systems that handle nuance, integrate diverse data streams, scale effectively, enable computational exploration, and enforce rigorous security.</div><div>Although the Digital Imaging and Communications in Medicine (DICOM) standard revolutionized radiology, it is imperative to explore its adequacy in addressing modern digital pathology's orchestration needs. Designed more than 30 years ago, DICOM reflects assumptions and architectural choices that predate modular software, cloud computing, and AI-driven workflows.</div><div>This article shows that by embedding metadata, annotations, and communication protocols into a unified container, DICOM limits interoperability and exposes architectural vulnerabilities. The article begins by examining these innate design risks, then challenges DICOM's interoperability claims, and ultimately presents a modular, standards-aligned alternative.</div><div>The article argues that separating image data from orchestration logic improves scalability, security, and performance. Standards such as HL7 FHIR (Health Level Seven Fast Healthcare Interoperability Resources) and modern databases manage clinical metadata; formats like Scalable Vector Graphics handle annotations; and fast, cloud-native file transfer protocols, and microservices support tile-level image access. This separation of concerns allows each component to evolve independently, optimizes performance across the system, and better adapts to emerging AI-driven workflows—capabilities that are inherently constrained in monolithic architectures where these elements are tightly coupled.</div><div>It further shows that security requirements should not be embedded within the DICOM standard itself. Instead, security must be addressed through a layered, format-independent framework that spans systems, networks, applications, and data governance. Security is not a discrete feature but an overarching discipline—defined by its own evolving set of standards and best practices. Overlays such as those outlined in the National Institute of Standards and Technology SP 800-53 support modern Transport Layer Security, single sign-on, cryptographic hashing, and other controls that protect data streams without imposing architectural constraints or restricting technological choices.</div><div>Pathology stands at a rare inflection point. Unlike radiology, where DICOM is deeply entrenched, pathology workflows still operate in polyglot environments—leveraging proprietary formats, hybrid standards, and emerging cloud-native tools. This diversity, often seen as a limitation, offers a clean slate: an opportunity to architect a modern, modular infrastructure free from legacy constraints. While a full departure from DICOM is unnecessary, pathology is uniquely position","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100450"},"PeriodicalIF":0.0,"publicationDate":"2025-08-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144925135","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-08-01DOI: 10.1016/j.jpi.2025.100444
Ilgar I. Guseinov, Arnab Bhowmik, Somaia AbuBaker, Anna E. Schmaus-Klughammer, Thomas Spittler
The rapid advancements in digital pathology, particularly in whole-slide imaging (WSI), have transformed remote histological analysis by enabling high-resolution digitization and real-time consultations. However, these workflows place significant demands on network infrastructure, requiring high bandwidth, low latency, and consistent performance. Whereas 5G networks have been extensively studied in controlled lab environments, their real-world applications in clinical settings remain underexplored.
This study provides a comparative analysis of 5G Campus Networks (5G CN) and traditional institutional networks, focusing on their performance during remote pathology tasks. Key metrics such as throughput, latency, and image quality were evaluated under various device loads to simulate real-world conditions. Although 5G CN did not consistently outperform in absolute throughput, it demonstrated superior adaptability, lower latency, and reduced variability, ensuring stable performance even with increasing network demand. These attributes are critical for time-sensitive workflows like frozen section analysis, where reliability and speed are paramount.
The findings highlight the potential of 5G CN to support emerging digital pathology applications, including real-time consultation. Furthermore, the study underscores the need for future research on 5G slicing and its ability to optimize network resources for high-demand medical applications. This work provides valuable insights into optimizing network infrastructure for the evolving demands of remote diagnostics in digital pathology.
{"title":"Comparative analysis of a 5G campus network and existing networks for real-time consultation in remote pathology","authors":"Ilgar I. Guseinov, Arnab Bhowmik, Somaia AbuBaker, Anna E. Schmaus-Klughammer, Thomas Spittler","doi":"10.1016/j.jpi.2025.100444","DOIUrl":"10.1016/j.jpi.2025.100444","url":null,"abstract":"<div><div>The rapid advancements in digital pathology, particularly in whole-slide imaging (WSI), have transformed remote histological analysis by enabling high-resolution digitization and real-time consultations. However, these workflows place significant demands on network infrastructure, requiring high bandwidth, low latency, and consistent performance. Whereas 5G networks have been extensively studied in controlled lab environments, their real-world applications in clinical settings remain underexplored.</div><div>This study provides a comparative analysis of 5G Campus Networks (5G CN) and traditional institutional networks, focusing on their performance during remote pathology tasks. Key metrics such as throughput, latency, and image quality were evaluated under various device loads to simulate real-world conditions. Although 5G CN did not consistently outperform in absolute throughput, it demonstrated superior adaptability, lower latency, and reduced variability, ensuring stable performance even with increasing network demand. These attributes are critical for time-sensitive workflows like frozen section analysis, where reliability and speed are paramount.</div><div>The findings highlight the potential of 5G CN to support emerging digital pathology applications, including real-time consultation. Furthermore, the study underscores the need for future research on 5G slicing and its ability to optimize network resources for high-demand medical applications. This work provides valuable insights into optimizing network infrastructure for the evolving demands of remote diagnostics in digital pathology.</div></div>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100444"},"PeriodicalIF":0.0,"publicationDate":"2025-08-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144925495","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The increasing availability of biomedical data is helping to design more robust deep learning (DL) algorithms to analyze biomedical samples. Currently, one of the main limitations to training DL algorithms to perform a specific task is the need for medical experts to manually label the data. Automatic methods to label data exist; however, automatic labels can be noisy, and it is not completely clear in which situations they can be used to train DL models.
This paper aims to investigate under which circumstances automatic labels can be used to train a DL model for the classification of whole slide images. The analysis involves multiple architectures, such as convolutional neural networks and vision transformer, and 10,604 WSIs as training data, collected from three use cases: celiac disease, lung cancer, and colon cancer, which include respectively binary, multiclass, and multilabel data. The results identify 10% as the percentage of noisy labels before a performance drop-off, so to train effective models for the classification of WSIs, reaching, respectively, F1-scores of 0.906, 0.757, and 0.833. Therefore, an algorithm generating automatic labels needs to stay within this range to be adopted, as shown by the application of Semantic Knowledge Extractor Tool as a tool to automatically extract concepts and use them as labels. Automatic labels are as effective as manual labels in this case, achieving solid performance comparable to that obtained by training models with manual labels.
{"title":"Automatic labels are as effective as manual labels in digital pathology images classification with deep learning","authors":"Niccolo Marini , Stefano Marchesin , Lluis Borras Ferris , Simon Püttmann , Marek Wodzinski , Riccardo Fratti , Damian Podareanu , Alessandro Caputo , Svetla Boytcheva , Simona Vatrano , Filippo Fraggetta , Iris Nagtegaal , Gianmaria Silvello , Manfredo Atzori , Henning Müller","doi":"10.1016/j.jpi.2025.100462","DOIUrl":"10.1016/j.jpi.2025.100462","url":null,"abstract":"<div><div>The increasing availability of biomedical data is helping to design more robust deep learning (DL) algorithms to analyze biomedical samples. Currently, one of the main limitations to training DL algorithms to perform a specific task is the need for medical experts to manually label the data. Automatic methods to label data exist; however, automatic labels can be noisy, and it is not completely clear in which situations they can be used to train DL models.</div><div>This paper aims to investigate under which circumstances automatic labels can be used to train a DL model for the classification of whole slide images. The analysis involves multiple architectures, such as convolutional neural networks and vision transformer, and 10,604 WSIs as training data, collected from three use cases: celiac disease, lung cancer, and colon cancer, which include respectively binary, multiclass, and multilabel data. The results identify 10% as the percentage of noisy labels before a performance drop-off, so to train effective models for the classification of WSIs, reaching, respectively, F1-scores of 0.906, 0.757, and 0.833. Therefore, an algorithm generating automatic labels needs to stay within this range to be adopted, as shown by the application of Semantic Knowledge Extractor Tool as a tool to automatically extract concepts and use them as labels. Automatic labels are as effective as manual labels in this case, achieving solid performance comparable to that obtained by training models with manual labels.</div></div>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100462"},"PeriodicalIF":0.0,"publicationDate":"2025-08-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144841531","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-08-01DOI: 10.1016/j.jpi.2025.100446
Orly Ardon , Allyne Manzo , Jamaal Spencer , Victor E. Reuter , Meera Hameed , Matthew G. Hanna
<div><h3>Background</h3><div>Digital pathology requires additional resources such as specialized whole slide imaging systems, staffing, space, and information technology infrastructure. Optimization of slide scanner throughput and quality are critical to achieve proper digital scanning operations. However, vendor supplied scanner throughput and scan speeds are often cited for a theoretical 15 × 15 mm tissue area and do not capture the real-world complexities of pathology slides or clinical workflows that contribute to the total time to scan a glass slide (e.g., scanner operator time). This study compares real-world scanner throughput using clinically generated glass slides, evaluating image quality errors, and total true scan time for seven different vendors' commercially available high-throughput scanners.</div></div><div><h3>Design</h3><div>Glass slides generated in a tertiary care CLIA-certified lab were retrieved from the departmental slide library including biopsies, surgical resections, and departmental consultation material from all surgical pathology subspecialties. Glass slide stain types include hematoxylin and eosin, immunohistochemical stains, or special stains per routine lab protocols. Slides were sequentially scanned by digital scan technicians on 16 different whole slide scanners from 7 different hardware vendor manufacturers. Two senior digital scan technicians reviewed each digital image that was generated from this study. One pathologist reviewed the set of slides for missing tissue determination. Scan times including scanner scan time, and time dedicated for pre- and post-scan work were recorded and summarized for the slide set for each scanner. Whole slide scanner models used in this study included: Leica Aperio AT2 and GT450 (Leica Biosystems, Buffalo Grove, Illinois); 3DHistech Pannoramic 1000, Philips UFS (Philips, Amsterdam, the Netherlands); Hamamatsu NanoZoomer S360 (Hamamatsu, Japan), Hologic Genius (Marlborough, MA), Huron TissueScope iQ (St. Jacobs Ontario, Canada) and 2-head Pramana Spectral HT scanning system (Pramana, Inc., Cambridge MA). Scanning was performed at ×40 equivalent magnification (∼0.25 μm per pixel) on each device, except for the Aperio AT2 and Huron TissueScope iQ which was ×20 equivalent magnification (0.5 μm per pixel). All scanner data were anonymized to guarantee unbiased interpretation of the results.</div></div><div><h3>Results</h3><div>347 glass slides representing real-world daily cases were assembled as a standardized slide set that was sequentially scanned on each device in this study. Variation in scan times for both the scanner model and labor time required to operate the scanner device were recorded. Actual instrument run time (e.g., scanner time) ranged between 7:30 and 43:02 (hours:minutes), the dedicated technician scanner operation time ranged from 1:30 to 9:24 h, and the total run time for each set, including the technician's time ranged from 13:30 to 47:02 h. Manual quality contro
{"title":"Digital slide scanning at scale: Comparison of whole slide imaging devices in a clinical setting","authors":"Orly Ardon , Allyne Manzo , Jamaal Spencer , Victor E. Reuter , Meera Hameed , Matthew G. Hanna","doi":"10.1016/j.jpi.2025.100446","DOIUrl":"10.1016/j.jpi.2025.100446","url":null,"abstract":"<div><h3>Background</h3><div>Digital pathology requires additional resources such as specialized whole slide imaging systems, staffing, space, and information technology infrastructure. Optimization of slide scanner throughput and quality are critical to achieve proper digital scanning operations. However, vendor supplied scanner throughput and scan speeds are often cited for a theoretical 15 × 15 mm tissue area and do not capture the real-world complexities of pathology slides or clinical workflows that contribute to the total time to scan a glass slide (e.g., scanner operator time). This study compares real-world scanner throughput using clinically generated glass slides, evaluating image quality errors, and total true scan time for seven different vendors' commercially available high-throughput scanners.</div></div><div><h3>Design</h3><div>Glass slides generated in a tertiary care CLIA-certified lab were retrieved from the departmental slide library including biopsies, surgical resections, and departmental consultation material from all surgical pathology subspecialties. Glass slide stain types include hematoxylin and eosin, immunohistochemical stains, or special stains per routine lab protocols. Slides were sequentially scanned by digital scan technicians on 16 different whole slide scanners from 7 different hardware vendor manufacturers. Two senior digital scan technicians reviewed each digital image that was generated from this study. One pathologist reviewed the set of slides for missing tissue determination. Scan times including scanner scan time, and time dedicated for pre- and post-scan work were recorded and summarized for the slide set for each scanner. Whole slide scanner models used in this study included: Leica Aperio AT2 and GT450 (Leica Biosystems, Buffalo Grove, Illinois); 3DHistech Pannoramic 1000, Philips UFS (Philips, Amsterdam, the Netherlands); Hamamatsu NanoZoomer S360 (Hamamatsu, Japan), Hologic Genius (Marlborough, MA), Huron TissueScope iQ (St. Jacobs Ontario, Canada) and 2-head Pramana Spectral HT scanning system (Pramana, Inc., Cambridge MA). Scanning was performed at ×40 equivalent magnification (∼0.25 μm per pixel) on each device, except for the Aperio AT2 and Huron TissueScope iQ which was ×20 equivalent magnification (0.5 μm per pixel). All scanner data were anonymized to guarantee unbiased interpretation of the results.</div></div><div><h3>Results</h3><div>347 glass slides representing real-world daily cases were assembled as a standardized slide set that was sequentially scanned on each device in this study. Variation in scan times for both the scanner model and labor time required to operate the scanner device were recorded. Actual instrument run time (e.g., scanner time) ranged between 7:30 and 43:02 (hours:minutes), the dedicated technician scanner operation time ranged from 1:30 to 9:24 h, and the total run time for each set, including the technician's time ranged from 13:30 to 47:02 h. Manual quality contro","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100446"},"PeriodicalIF":0.0,"publicationDate":"2025-08-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144925496","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-23eCollection Date: 2025-08-01DOI: 10.1016/j.jpi.2025.100464
Lydia A Schoenpflug, Ruben Bagan Benavides, Marta Nowak, Fahime Sheikhzadeh, Arash Moayyedi, Kamil Wasag, Jacob Reimers, Michael Zhou, Raghavan Venugopal, Bettina Sobottka, Yasmin Koeller, Michael Rivers, Holger Moch, Yao Nie, Viktor H Koelzer
Federated learning (FL) allows institutions to collaboratively train deep learning models while maintaining data privacy, a critical aspect in fields like computational pathology (CPATH). However, existing studies focus on performance improvement in simulated environments and overlook practical aspects of FL. In this study, we address this need by transparently sharing the challenges encountered in the real-world application of FL for a clinical CPATH use case. We set up a FL framework consisting of three clients and a central server to jointly train deep learning models for digital immune phenotyping in metastatic melanoma, utilizing the NVIDIA Federated Learning Application Runtime Environment (NVIDIA FLARE) across four separate networks from institutes in four countries. Our findings reveal several key challenges: First, the FL model performs the best across all clients' test sets but does not outperform all local models on their own client test set. Second, long experiment duration due to system and data heterogeneity limited experiment frequency, alleviated by optimizing local client epochs. Third, infrastructure design was hindered by hospital and corporate network restrictions, necessitating an open port for the server, which we resolved by deploying the server on an Amazon Web Services infrastructure within a semi-public network. Lastly, effective experiment management required IT expertise and strong familiarity with NVIDIA FLARE to enable orchestration, code management, parameter configuration, and logging. Our findings provide a practical perspective on implementing FL for CPATH, advocating for greater transparency in future research and the development of best practices and guidelines for implementing FL in real-world healthcare settings.
联邦学习（FL）允许机构在保持数据隐私的同时协同训练深度学习模型，这是计算病理学（CPATH）等领域的一个关键方面。然而，现有的研究侧重于模拟环境中的性能改进，而忽视了FL的实际应用。在本研究中，我们通过透明地分享FL在临床CPATH用例中的实际应用中遇到的挑战来解决这一需求。我们建立了一个由三个客户端和一个中央服务器组成的FL框架，利用NVIDIA联邦学习应用运行时环境（NVIDIA FLARE）在来自四个国家的机构的四个独立网络上联合训练用于转移性黑色素瘤数字免疫表型的深度学习模型。我们的发现揭示了几个关键的挑战：首先，FL模型在所有客户的测试集中表现最好，但在他们自己的客户测试集中表现不优于所有本地模型。其次，由于系统和数据的异构性，实验时间长，限制了实验频率，通过优化本地客户端时间可以缓解这一问题。第三，基础设施设计受到医院和企业网络限制的阻碍，需要为服务器提供一个开放端口，我们通过将服务器部署在半公共网络中的Amazon Web Services基础设施上来解决这个问题。最后，有效的实验管理需要IT专业知识和对NVIDIA FLARE的熟悉程度，以实现编排、代码管理、参数配置和日志记录。我们的研究结果为在CPATH中实施FL提供了一个实用的视角，提倡在未来的研究中提高透明度，并为在现实世界的医疗环境中实施FL制定最佳实践和指南。
{"title":"Navigating real-world challenges: A case study on federated learning in computational pathology.","authors":"Lydia A Schoenpflug, Ruben Bagan Benavides, Marta Nowak, Fahime Sheikhzadeh, Arash Moayyedi, Kamil Wasag, Jacob Reimers, Michael Zhou, Raghavan Venugopal, Bettina Sobottka, Yasmin Koeller, Michael Rivers, Holger Moch, Yao Nie, Viktor H Koelzer","doi":"10.1016/j.jpi.2025.100464","DOIUrl":"10.1016/j.jpi.2025.100464","url":null,"abstract":"<p><p>Federated learning (FL) allows institutions to collaboratively train deep learning models while maintaining data privacy, a critical aspect in fields like computational pathology (CPATH). However, existing studies focus on performance improvement in simulated environments and overlook practical aspects of FL. In this study, we address this need by transparently sharing the challenges encountered in the real-world application of FL for a clinical CPATH use case. We set up a FL framework consisting of three clients and a central server to jointly train deep learning models for digital immune phenotyping in metastatic melanoma, utilizing the NVIDIA Federated Learning Application Runtime Environment (NVIDIA FLARE) across four separate networks from institutes in four countries. Our findings reveal several key challenges: First, the FL model performs the best across all clients' test sets but does not outperform all local models on their own client test set. Second, long experiment duration due to system and data heterogeneity limited experiment frequency, alleviated by optimizing local client epochs. Third, infrastructure design was hindered by hospital and corporate network restrictions, necessitating an open port for the server, which we resolved by deploying the server on an Amazon Web Services infrastructure within a semi-public network. Lastly, effective experiment management required IT expertise and strong familiarity with NVIDIA FLARE to enable orchestration, code management, parameter configuration, and logging. Our findings provide a practical perspective on implementing FL for CPATH, advocating for greater transparency in future research and the development of best practices and guidelines for implementing FL in real-world healthcare settings.</p>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"100464"},"PeriodicalIF":0.0,"publicationDate":"2025-07-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12357140/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144875696","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-18eCollection Date: 2025-08-01DOI: 10.1016/j.jpi.2025.100463
Ekaterina Redekop, Mara Pleasure, Zichen Wang, Anthony Sisk, Yang Zong, Kimberly Flores, William Speier, Corey W Arnold
Histological analysis of biopsy samples by pathologists can require the evaluation of complex three-dimensional (3D) tissue structures. This process involves studying the same tissue region across slides, which requires laborious zooming and panning for localization. Additionally, standard deep learning frameworks typically focus on cross-sections cut from biopsy specimens, limiting their ability to capture 3D tissue spatial information. We present a novel framework that constructs 2.5D biopsy cores via the extraction and co-alignment of serial tissue sections using a novel morphology-preserving alignment framework. These 2.5D cores can then be used for enhanced viewing by pathologists and as input to video transformer models that can capture depth-wide spatial dependencies. We used our framework to construct 2.5D cores for 10,210 prostate biopsies, 156 breast biopsies, and 1869 renal biopsies. To evaluate the utility of the cores for downstream tasks, we performed additional studies in prostate cancer by: (1) training a deep learning-based cancer grading model and (2) conducting a reader study with pathologists.
{"title":"Generating 2.5D pathology for enhanced viewing and AI diagnosis.","authors":"Ekaterina Redekop, Mara Pleasure, Zichen Wang, Anthony Sisk, Yang Zong, Kimberly Flores, William Speier, Corey W Arnold","doi":"10.1016/j.jpi.2025.100463","DOIUrl":"10.1016/j.jpi.2025.100463","url":null,"abstract":"<p><p>Histological analysis of biopsy samples by pathologists can require the evaluation of complex three-dimensional (3D) tissue structures. This process involves studying the same tissue region across slides, which requires laborious zooming and panning for localization. Additionally, standard deep learning frameworks typically focus on cross-sections cut from biopsy specimens, limiting their ability to capture 3D tissue spatial information. We present a novel framework that constructs 2.5D biopsy cores via the extraction and co-alignment of serial tissue sections using a novel morphology-preserving alignment framework. These 2.5D cores can then be used for enhanced viewing by pathologists and as input to video transformer models that can capture depth-wide spatial dependencies. We used our framework to construct 2.5D cores for 10,210 prostate biopsies, 156 breast biopsies, and 1869 renal biopsies. To evaluate the utility of the cores for downstream tasks, we performed additional studies in prostate cancer by: (1) training a deep learning-based cancer grading model and (2) conducting a reader study with pathologists.</p>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"100463"},"PeriodicalIF":0.0,"publicationDate":"2025-07-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12355132/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144875695","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-09DOI: 10.1016/j.jpi.2025.100461
Ryan Erik Landvater, Michael David Olp, Mustafa Yousif, Ulysses Balis
A modern digital pathology vendor-agnostic binary slide format specifically targeting the unmet need of efficient real-time transfer and display has not yet been established. The growing adoption of digital pathology only intensifies the need for an intermediary digital slide format that emphasizes performance for use between slide servers and image management software. The DICOM standard is a well-established format widely used for the long-term storage of both images and associated critical metadata. However, it was inherently designed for radiology rather than digital pathology, a discipline that imposes a unique set of performance requirements due to high-speed multi-pyramidal rendering within whole slide viewer applications. Here, we introduce the Iris file extension, a binary container specification explicitly designed for performance-oriented whole slide image (WSI) viewer systems. The Iris file extension specification is explicit and straightforward, adding modern compression support, a dynamic structure with fully optional metadata features, computationally trivial deep file validation, corruption recovery capabilities, and slide annotations. In addition to the file specification document, we provide source code to allow for (de)serialization and validation of a binary stream against the standard. We also provide corresponding binary builds with C++, Python, and JavaScript language support. Finally, we provide full encoder and decoder implementation source code, as well as binary builds (part of the separate Iris Codec Community module), with language bindings for C++ and Python, allowing for easy integration with existing WSI solutions. We provide the Iris File Extension specification openly to the community in the form of a Creative Commons Attribution-No Derivative 4.0 International license.
{"title":"The Iris File Extension","authors":"Ryan Erik Landvater, Michael David Olp, Mustafa Yousif, Ulysses Balis","doi":"10.1016/j.jpi.2025.100461","DOIUrl":"10.1016/j.jpi.2025.100461","url":null,"abstract":"<div><div>A modern digital pathology vendor-agnostic binary slide format specifically targeting the unmet need of efficient real-time transfer and display has not yet been established. The growing adoption of digital pathology only intensifies the need for an intermediary digital slide format that emphasizes performance for use between slide servers and image management software. The DICOM standard is a well-established format widely used for the long-term storage of both images and associated critical metadata. However, it was inherently designed for radiology rather than digital pathology, a discipline that imposes a unique set of performance requirements due to high-speed multi-pyramidal rendering within whole slide viewer applications. Here, we introduce the Iris file extension, a binary container specification explicitly designed for performance-oriented whole slide image (WSI) viewer systems. The Iris file extension specification is explicit and straightforward, adding modern compression support, a dynamic structure with fully optional metadata features, computationally trivial deep file validation, corruption recovery capabilities, and slide annotations. In addition to the file specification document, we provide source code to allow for (de)serialization and validation of a binary stream against the standard. We also provide corresponding binary builds with C++, Python, and JavaScript language support. Finally, we provide full encoder and decoder implementation source code, as well as binary builds (part of the separate Iris Codec Community module), with language bindings for C++ and Python, allowing for easy integration with existing WSI solutions. We provide the Iris File Extension specification openly to the community in the form of a Creative Commons Attribution-No Derivative 4.0 International license.</div></div>","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100461"},"PeriodicalIF":0.0,"publicationDate":"2025-07-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144696900","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-07DOI: 10.1016/j.jpi.2025.100457
Elzbieta Budginaite , Derek R. Magee , Maximilian Kloft , Henry C. Woodruff , Heike I. Grabsch
{"title":"Corrigendum to “Computational methods for metastasis detection in lymph nodes and characterization of the metastasis-free lymph node microarchitecture: A systematic-narrative hybrid review”. Journal of Pathology Informatics 15(2024) 100367","authors":"Elzbieta Budginaite , Derek R. Magee , Maximilian Kloft , Henry C. Woodruff , Heike I. Grabsch","doi":"10.1016/j.jpi.2025.100457","DOIUrl":"10.1016/j.jpi.2025.100457","url":null,"abstract":"","PeriodicalId":37769,"journal":{"name":"Journal of Pathology Informatics","volume":"18 ","pages":"Article 100457"},"PeriodicalIF":0.0,"publicationDate":"2025-07-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144570460","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号