This Week in Pathology — Sep 5, 2026

Welcome to This Week in Pathology. This week we're covering 10 notable papers spanning three broad themes: artificial intelligence and quantification in the laboratory, consensus and standardization efforts that shape how we report, and the molecular reclassification of morphologically ambiguous tumours. Let's dive in.

We start with artificial intelligence, where the conversation has shifted from can it work to how do we run it safely. Rakha and colleagues, writing in Virchows Archiv, offer a practical framework for laboratories adopting commercially available, regulatory-approved algorithms rather than building their own [2]. Their central argument is that regulatory approval is the start of the process, not the end. They separate algorithm validation from local verification and from continuous assurance, and they insist that these are different tasks requiring different work, with the intensity scaled to what the tool actually does. A diagnostic algorithm, a biomarker quantifier, a workflow triage tool and a generative text tool carry very different risks and should not be governed identically. The practical recommendations cover interoperability, human oversight, user competency, performance monitoring, incident reporting and, importantly, proportionate re-verification after software updates, all embedded in the quality management system you already run. If your department is signing an artificial intelligence contract this year, this is the document to hand to your quality lead.

That governance question gets sharper when you look at what artificial intelligence actually measures. In Histopathology, Nazli and colleagues asked ten experienced liver pathologists from four institutions to assess steatosis on twenty whole-slide images from metabolic dysfunction-associated steatotic liver disease cases, using four different methods [5]. The pathologists differed substantially in terminology, in how they defined droplet size, and in their quantification approach. Most strikingly, visual estimates exceeded the artificial intelligence measurements by roughly one and a half to three and a half times, consistently in the same direction. The choice of method was the main driver of disagreement: steatosis proportionate area, meaning the fraction of tissue occupied by fat, is conceptually different from the percentage of hepatocytes containing fat, and using them interchangeably amplifies discrepancy, especially in cases rich in small droplets. The authors derived calibration equations to convert between visual and digital scales, but they flag these as exploratory and requiring independent validation. The immediate message is more modest and more useful: state which method you used when you report steatosis, because the number is meaningless without it. Alongside this, Dedic and colleagues in Virchows Archiv review deep learning for axillary lymph node assessment in breast cancer, and draw a clear line between two applications [8]. Detecting metastases on lymph node slides is essentially solved, with near-perfect sensitivity, already entering clinical use, and reducing immunohistochemistry workload and review time without loss of accuracy. Predicting nodal status from the primary tumour is not there: external performance ranged from no better than chance up to reasonable discrimination, dropped across institutions, and studies often failed to state whether isolated tumour cells or micrometastases counted as a positive node. Inconsistent endpoints make the literature hard to pool.

Our second theme is standardization and the unglamorous determinants of diagnostic quality. The International Society of Urological Pathology convened in Vienna in September 2025, and Working Group 2, reported by Downes and colleagues in the American Journal of Surgical Pathology, tackled treatment effects in the urinary bladder [1]. A premeeting survey of the membership exposed real uncertainty about tumour regression grading and wide variability in how post-resection and post-intravesical-therapy cases are reported. Consensus was reached on 18 of 19 statements, supporting standardized terminology for post-therapy reporting, interdisciplinary development of a bladder tumour regression grading scheme, and selective use of molecular and immunohistochemical testing after treatment. The one statement that failed to reach consensus is telling: using molecular tests to separate benign mimickers from carcinoma. That remains contested. Complementing this from the other end of the workflow, Vaz Alvares Fernandes reviews preanalytical variables in the American Journal of Clinical Pathology, and makes the case that a specimen's diagnostic value is largely determined before anyone looks down a microscope [4]. Ischaemia time, delayed or inadequate fixation, tissue thickness, transport and labelling problems, orientation, grossing variability and tissue contamination all degrade morphology, margins, staging and biomarker interpretation. The stakes are highest in small biopsies and precision oncology specimens, where the same limited tissue must support diagnosis, immunohistochemistry, fluorescence in situ hybridization, polymerase chain reaction testing and next-generation sequencing. Gross examination is singled out as the pathology-controlled step that matters most, because block selection sets the ceiling on everything downstream. The framing is that preanalytical control is a patient safety process, not technical background noise. In the same standardization vein, Law and Choi review inflammatory bowel disease-associated colorectal dysplasia in Virchows Archiv, mapping the expanding morphologic landscape against the sixth edition of the World Health Organization classification [3]. The less well recognized variants are the problem: subtle cytologic atypia overlapping with regenerative change, frequently misclassified as reactive or indefinite for dysplasia, and often endoscopically invisible or flat. That matters because many of these lesions are associated with advanced neoplasia on follow-up and with synchronous or metachronous neoplasia in the same colonic segment. The review provides a standardized reporting framework, including how to handle lesions with partial features that fall short of definitive dysplasia.

Our third theme is molecular and immunohistochemical reclassification of tumours that look like something else. Two papers this week expand aggressive renal cell carcinoma phenotypes. In the American Journal of Surgical Pathology, Kandukuri and colleagues describe five succinate dehydrogenase A-deficient renal cell carcinomas [9]. Only one was correctly called at the outset; the others were signed out as collecting duct carcinoma or high-grade carcinoma not otherwise specified, with sequencing making the diagnosis. Papillary and nested architecture, eosinophilic cytoplasm and cytoplasmic vacuoles with inclusions were common. The critical practical point is that all five lost succinate dehydrogenase B expression, but succinate dehydrogenase A staining was retained in one case despite a confirmed mutation, so immunohistochemistry alone can mislead and sequencing is recommended whenever succinate dehydrogenase B loss is seen. Unlike most succinate dehydrogenase B-deficient tumours, these behaved aggressively, with two of four patients developing metastases within roughly one to three years. Running in parallel, Wang and colleagues in Virchows Archiv report six high-grade papillary-patterned renal cell carcinomas carrying folliculin frameshift mutations [10]. Folliculin-mutated tumours are supposed to be low-grade, eosinophilic and indolent; these were infiltrative, necrotic and high-grade, overlapping morphologically with papillary carcinoma, fumarate hydratase-deficient carcinoma and rearranged carcinomas. All strongly expressed GPNMB, most mutations were germline, and half of the patients developed distant metastases, with two dying of disease. The lesson from both papers is the same: in high-grade papillary-patterned kidney tumours that resist classification, sequencing changes both the diagnosis and the family's risk. Two further biomarker papers round this out. In the American Journal of Surgical Pathology, Feng and colleagues examined folate receptor alpha across more than four hundred tubal epithelial samples [7]. High expression was present in roughly three quarters of serous tubal intraepithelial carcinomas and of high-grade serous carcinomas, with no significant difference between them, while serous tubal intraepithelial lesions and earlier changes were predominantly low or negative. Interpreted alongside morphology, p53 and Ki-67, folate receptor alpha may add confidence along the difficult intraepithelial lesion to intraepithelial carcinoma boundary. And in Histopathology, Alnaqshabandi and colleagues showed that commercially available low-risk human papillomavirus chromogenic in situ hybridization was positive in all twelve digital papillary adenocarcinomas and negative in all eight tubular adenomas, with the reverse pattern for BRAF V600E [6]. The catch is that the signal is punctate and nuclear, sometimes visible only at high magnification, so a pathologist expecting diffuse staining may call it negative.

If you only have time for one paper this week, make it the Virchows Archiv framework for laboratory implementation and continuous assurance of artificial intelligence [2]. Nearly every department will be asked to adopt an approved algorithm in the next few years, and this is the first practical guidance on what the laboratory itself is responsible for once the vendor's validation ends.

Here are the key takeaways from this week in Pathology. First, regulatory approval of an artificial intelligence tool does not discharge your local obligation to verify, monitor and re-verify it after updates. Second, when you report steatosis, name your method, because proportionate area and percentage of hepatocytes containing fat are different measures and visual estimates run systematically higher than machine measurement. Third, deep learning for lymph node metastasis detection is clinically ready; predicting nodal status from primary tumour morphology is not. Fourth, in high-grade renal cell carcinoma with papillary architecture that does not fit a recognized subtype, send it for sequencing, because succinate dehydrogenase A immunohistochemistry can be falsely retained and folliculin mutation carries germline implications. And fifth, know your staining patterns: punctate nuclear low-risk human papillomavirus signal in digital papillary adenocarcinoma is easy to miss if you expect a diffuse blush.

That's your roundup for This Week in Pathology. The full transcript and references are available on the episode page in your AudioScholar library. This is an AI-curated summary — for clinical decisions, always consult primary sources and current guidelines. See you next week.

If this weekly briefing is useful, follow the show in your podcast app so new episodes arrive automatically. And if it saves you time, a quick rating in your podcast app helps other physicians find the show.

This is an automated summary generated by artificial intelligence, which can make mistakes. Always review the original source materials.