Artificial Intelligence for Diagnostic Pathology in Nephrology

The study “Computational Characterization of Lymphocyte Topology on Whole Slide Images of Glomerular Diseases,”¹ published in this issue of JASN, presents an ambitious and carefully designed computational pathology study aimed at characterizing lymphocyte spatial organization in glomerular diseases. Using whole slide pathology images from two clinical cohorts, the NEPhrotic syndrome sTUdy NEtwork (NEPTUNE) and Cure Glomerulonephropathy (CureGN) cohorts, the authors develop a multistep pipeline. Their computational method identifies lymphocytes in images, constructs cell graphs, defines dense and sparse “immune habitats,” and extracts a number of “graph topology features” describing connectivity, clustering, and structural cohesion. These features are then linked to clinical outcomes, specifically a relevant decline in eGFR or kidney failure.

Several findings stand out. First, from a technical standpoint, it is noteworthy that many of the topological features retain prognostic value even after adjustment for standard clinical covariates, suggesting that spatial immune architecture contains clinically relevant information that is not captured by conventional pathology or by simple lymphocyte density measures. This is very interesting because it provides a clear justification of using artificial intelligence (AI)–based assessment of pathology slides, as opposed to simple metrics such as cell counts. Second, the authors show that “sparse” and “dense” habitats contribute to clinical outcomes in distinct ways, where increased “connectivity” in dense regions appears associated with more severe inflammatory states. This finding is relevant because it shows that computationally defined patterns are interpretable in biologic terms and can provide new hypotheses in how we look at biologic processes, such as inflammation in the kidney. Third, the authors show that external validation, which in this case means training on NEPTUNE and testing on CureGN, reveals meaningful differences between cohorts. This phenomenon is called batch effect and is an important reminder that any image analysis model should always be tested on external datasets and not just on a randomly held-out set from the training cohort, otherwise the performance will be inflated. This is also defined in well-established quality criteria for biomarkers in medical research, such as the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis guidance, where external validation constitutes the highest evidence grade.²

Taken together, the study at hand makes two main contributions: it offers a biologically interpretable framework for understanding tissue topology in kidney disease, and it provides a prognostic model with respectable discrimination across independent datasets. For nephrology, where histologic biomarkers for progression exist but the clinical implementation of computational biomarkers has not fully taken off, this work supports and expands previous work such as recent work by Hölscher et al.³ There is now ample evidence that computational pathology should be considered an important candidate for future biomarkers in nephrology.⁴

It is valuable to situate the study within the broader landscape of AI today. Across many fields in machine learning, we have repeatedly observed that performance tends to improve as human priors are removed and as AI systems have more degrees of freedom to learn patterns from data. An example from natural language processing is perhaps the clearest parallel. For many years, linguistic experts attempted to encode grammatical structures in computational workflows through handcrafted rules, motivated by decades of expert knowledge and theoretical insight. Yet the most effective AI systems today are large language models, which are trained end-to-end on raw text from the open internet. These models were not given grammatical rules; instead, they rediscovered the structure of language by optimizing predictive performance.

A similar pattern is now apparent in AI in medicine, and computational pathology in particular. Traditional computational pipelines rely on a predefined sequence of distinct steps such as segmentation of cells or tissue structures and extraction of hand-crafted visual features. Such features include cell counts, morphologic measurements, and graph descriptors and have been valuable for biologic insight. However, when the goal is predictive performance for clinical end points, these are often outperformed by end-to-end deep learning systems, which process raw images directly to predict outcomes without intermediate human-defined steps.^5,6 Several examples of such end-to-end workflows have already entered routine clinical deployment. Multiple companies have developed and brought to the market AI methods for predicting patient outcomes from pathology images using end-to-end AI methods that operate directly on whole slide images without explicit modeling of individual cell types.^7–9 These models work because they automatically discover hierarchical representations in visual data: recognizing cell types, spatial relationships, tissue-level patterns, and also latent morphologic signals that may not necessarily align with human-defined (or even human understandable) features.

Within this context, this study is both timely and complementary. The authors' graph topology approach provides interpretability and mechanistic grounding, which is not a given in end-to-end systems. It allows researchers to ask biologically meaningful questions, such as whether the structure of sparse inflammatory activity is an early marker of disease progression or whether dense lymphocyte connectivity reflects active fibrogenic processes. These are insights that end-to-end models implicitly use but do not explicitly reveal. However, for clinical deployment of a biomarker, an end-to-end foundation model–based approach could further improve performance.

Beyond a technical analysis of this study, it is important to ask: How can this method be translated to clinical routine, and what additional evidence is needed to safely do so? Adjacent areas of medicine, such as oncology, can provide useful examples on how to move from purely technical image analysis studies and toward clinical implementation of AI biomarkers. The European Society for Medical Oncology (ESMO) has recently introduced the first guidance for evaluating and classifying AI-derived biomarkers, ESMO basic requirements for AI-based biomarkers in oncology (EBAI).¹⁰ This framework provides expectations for analytical validity, clinical validity, and clinical utility. Under this framework, a pathology-based prognostic system such as the one proposed here would fall into the category of a “class C” biomarker, meaning that it is a “novel” biomarker. EBAI recommends the following validation: For class C1 (prognostic) biomarkers, the framework mandates validation using high-quality retrospective real-world data or clinical trial data. Crucially, it requires a rigorous comparison against multiple established prognostic markers used in routine clinical care to demonstrate that the AI tool provides independent value beyond the current standard. Furthermore, EBAI emphasizes that generalizability must be demonstrated across independent patient cohorts to ensure robustness against variations in data acquisition and population characteristics. This study aligns well with these emerging standards and ticks some of these boxes. However, if these guidelines from oncology biomarkers are taken as a reference, some additional points would have to be fulfilled, such as a rigorous head-to-head comparison, demonstrating that the AI model adds independent prognostic value beyond the current standard-of-care clinical risk models in multiple validation cohorts. In addition, professional societies could adapt ESMO's EBAI guidance to the specific clinical needs of kidney disease. Ultimately, such frameworks help ensure that promising algorithms can safely mature into tools that influence clinical decisions in real patient care.

In summary, this study demonstrates how computational pathology can reveal inflammatory microarchitectures that are not visible through traditional histology, and it opens new ways for mechanistic inquiry in glomerular disease. At the same time, it points toward a future in which nephrology could embrace even larger datasets and even more powerful AI models and adopt formal guidelines for AI biomarker development.

Supplementary Material

jasn-37-439-s001.pdf ^{(1.4MB, pdf)}

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/F598.

Author Contributions

Conceptualization: Jakob Nikolas Kather.

Methodology: Jakob Nikolas Kather.

Project administration: Jakob Nikolas Kather.

Resources: Jakob Nikolas Kather.

Software: Jakob Nikolas Kather.

Supervision: Jakob Nikolas Kather.

Writing – original draft: Jakob Nikolas Kather.

Writing – review & editing: Jakob Nikolas Kather.

Funding

None.