Artificial intelligence-powered digital pathology: A potential gold standard for kidney biopsy interpretation?

INTRODUCTION

Kidney diseases represent a significant global health burden, with an estimated 850 million people affected worldwide,^[1] contributing to substantial morbidity, mortality, and healthcare costs. Accurate and timely diagnosis is paramount for effective management and prognostication in nephrology. For decades, renal biopsy has served as the indispensable gold standard, providing critical histological insights into the underlying pathology of various kidney disorders. Pathologists meticulously examine tissue samples under light microscopy, often supplemented by immunofluorescence and electron microscopy, to identify specific lesions, assess disease activity and chronicity, and guide therapeutic decisions.^[2,3]

Despite its pivotal role, traditional renal pathology faces inherent challenges that can impact diagnostic consistency and reproducibility. The subjective nature of microscopic interpretation, coupled with inter-observer variability among pathologists, can lead to discrepancies in scoring and diagnosis.^[4,5] This variability can have profound implications for patient care, potentially delaying appropriate treatment or influencing clinical research outcomes. Furthermore, the increasing complexity of renal pathologies and the growing demand for specialized expertise highlight the urgent need for innovative approaches to enhance diagnostic precision and efficiency.

The rapid evolution of digital pathology, which involves the digitization of glass slides into whole-slide images (WSIs), has laid the groundwork for a transformative shift in pathological diagnosis. This technological advancement has paved the way for the integration of artificial intelligence (AI) into routine pathology workflows. AI, particularly deep learning (DL), a subset of machine learning, has demonstrated remarkable capabilities in image analysis, pattern recognition, and predictive modeling across various medical disciplines.^[6,7] In renal pathology, AI holds immense promise to overcome existing limitations by providing objective, quantitative, and standardized assessments of kidney biopsy specimens.

This review aims to provide a comprehensive overview of how AI is poised to revolutionize renal pathology, potentially establishing a new gold standard for kidney biopsy interpretation. We will explore the current state-of-theart applications of AI across the entire diagnostic pipeline, from ensuring sample adequacy and quantifying histological features to improving the reproducibility of established scoring systems. We will also discuss the exciting potential of AI in discovering novel histologic patterns that correlate with clinical outcomes, thereby advancing personalized medicine in nephrology. By synthesizing insights from pathology, nephrology, and data science, this article seeks to serve as a definitive guide for clinicians and researchers, fostering interdisciplinary collaboration and accelerating the adoption of AI-powered solutions in renal pathology practice and research.

PRACTICAL EXAMPLES OF AI IN RENAL PATHOLOGY

The application of AI in renal pathology has moved beyond theoretical concepts to deliver tangible results in both research and clinical practice. For instance, a multiclass segmentation model developed in 2019 demonstrated the ability to differentiate various kidney structures with a high degree of accuracy (weighted mean Dice coefficient 0.80–0.84), showing strong correlation with expert pathologist assessments.^[8] Another practical example is the use of AI for the quantitative assessment of interstitial fibrosis and tubular atrophy (IFTA), where AI algorithms have been shown to outperform human pathologists in terms of reproducibility.^[9] In the context of kidney transplantation, AI-powered tools are being developed to automate the Banff classification of allograft pathology, with initiatives like the MONKEY challenge fostering community-based development of models for inflammatory cell detection.^[10] Furthermore, in a study on immunoglobulin A (IgA) nephropathy, an AI model demonstrated an Area Under the Curve of 0.893 for predicting kidney outcomes when combining clinical and image features, showcasing the potential for enhanced prognostic prediction.^[11] These examples illustrate the concrete benefits of AI in improving the accuracy, reproducibility, and predictive power of renal pathology.

AI FOR STANDARDIZING SAMPLE ADEQUACY AND QUANTIFYING HISTOLOGICAL FEATURES

One of the initial critical steps in renal biopsy interpretation is ensuring the adequacy of the tissue sample. Insufficient tissue can lead to misdiagnosis or an inability to render a definitive diagnosis, necessitating repeat biopsies, which carry risks for the patient. AI-powered solutions are emerging to address this challenge by providing objective and real-time assessment of sample adequacy. These systems can analyze images of biopsy cores, even those captured by smartphone cameras at the bedside, to quantify the number of glomeruli, the length of cortical tissue, and the presence of other critical structures, thereby standardizing the initial quality control process.^[12,13] This capability is particularly valuable in resource-limited settings or for rapid assessment during biopsy procedures.

Beyond sample adequacy, AI excels at the precise and quantitative assessment of various histological features, which are often subject to inter-observer variability in traditional manual evaluation. Key pathological changes in kidney diseases, such as glomerulosclerosis, interstitial fibrosis, and tubular atrophy, are crucial for diagnosis, prognosis, and treatment monitoring. AI algorithms, particularly convolutional neural networks, have demonstrated remarkable accuracy in segmenting and quantifying these features from WSIs^[14,15] [Table 1 and Figure 1].

Close

Figure 1:

Proportion of primary artificial intelligence methods used in renal pathology.

Export to PPT

GLOMERULOSCLEROSIS QUANTIFICATION

Glomerulosclerosis, the scarring of glomeruli, is a common finding in many chronic kidney diseases and is a strong predictor of renal function decline. Conventionally, pathologists estimate the percentage of sclerosed glomeruli, which can be subjective. AI models have been developed to automatically detect and classify glomeruli as sclerotic or non-sclerotic with high precision. These models can quantify the extent of global and segmental glomerulosclerosis, providing an objective and reproducible measure that correlates well with clinical outcomes.^[16,17] The ability of AI to rapidly process large numbers of glomeruli across an entire WSI offers a significant advantage over manual counting, which is time-consuming and prone to sampling bias.

IFTA ASSESSMENT

IFTA is a critical marker of chronic kidney injury and progression to end-stage renal disease. The accurate quantification of IFTA is essential for prognostication and guiding therapeutic interventions. AI algorithms can precisely delineate areas of fibrosis and atrophy within the renal cortex, often outperforming human pathologists in terms of reproducibility and consistency.^[18-20] By leveraging advanced image analysis techniques, AI can provide a continuous, quantitative score for IFTA, moving beyond the semi-quantitative grading systems currently in use. This objective measurement can lead to more precise patient stratification and a better understanding of disease progression.

OTHER HISTOLOGICAL FEATURES

AI is also being applied to quantify other important histological features, including inflammation (e.g., tubulointerstitial inflammation, glomerulitis, peritubular capillaritis), vascular changes (e.g., arteriosclerosis, arteriolosclerosis), and specific lesion types (e.g., crescent formation, immune complex deposits). The ability of AI to accurately identify and quantify these diverse features across the entire biopsy specimen provides a comprehensive and objective pathological profile, which can be invaluable for both clinical practice and research.^[15,20,21] The integration of these quantitative AI assessments into routine diagnostic reports has the potential to enhance diagnostic accuracy, reduce inter-observer variability, and provide more granular information for personalized patient management.

IMPROVING REPRODUCIBILITY OF SCORING SYSTEMS: BANFF AND MESANGIAL HYPERCELLULARITY, ENDOCAPILLARY HYPERCELLULARITY, SEGMENTAL GLOMERULOSCLEROSIS, TUBULAR ATROPHY/INTERSTITIAL FIBROSIS, AND CRESCENTS (MEST-C)

Standardized scoring systems are crucial for consistent diagnosis, prognostication, and guiding treatment in renal pathology. However, even widely adopted systems such as the Banff classification for kidney allograft pathology and the MEST-C score for IgA nephropathy (IgAN) often suffer from inter-observer variability, limiting their full clinical utility. AI offers a powerful solution to enhance the reproducibility and objectivity of these complex scoring systems [Table 2].

AI IN BANFF CLASSIFICATION

The Banff classification is the international standard for diagnosing and grading kidney transplant rejection and other pathologies. It relies on the semi-quantitative assessment of various histological lesions, such as glomerulitis (g), peritubular capillaritis (ptc), interstitial inflammation (i), tubulitis (t), intimal arteritis (v), and IFTA (ci, ct). Despite detailed guidelines, significant inter-observer variability among pathologists in scoring these lesions has been well-documented, impacting diagnostic consistency and clinical decision-making.^[23,24]

AI, particularly DL models trained on large datasets of annotated kidney transplant biopsies, can automate the scoring of individual Banff lesions with high accuracy and reproducibility. These models can precisely segment and quantify inflammatory cells, assess the severity of tubulitis, and measure the extent of vascular changes, providing objective metrics that reduce subjective interpretation.^[8,25] By integrating these AI-derived scores, a more consistent and automated Banff classification system can be achieved, leading to improved diagnostic agreement among pathologists and more reliable prognostication for transplant recipients. Efforts by groups like the Banff Digital Pathology Working Group are actively promoting the development and validation of such AI tools to facilitate their clinical adoption.^[2]

AI IN MEST-C SCORING FOR IGAN

IgAN is the most common primary glomerulonephritis worldwide, and its prognosis is highly variable. The Oxford Classification of IgAN, which includes the MEST-C score, provides a standardized framework for assessing disease severity and predicting renal outcomes. Similar to Banff, the MEST-C scoring can be prone to inter-observer variability, particularly for features such as mesangial and endocapillary hypercellularity.^[26,27]

AI algorithms are being developed to objectively quantify each component of the MEST-C score. For instance, AI can accurately measure mesangial cellularity, detect and quantify endocapillary proliferation, and identify segmental glomerulosclerosis and crescents with high precision.^[28,29] By providing quantitative and reproducible scores for each MEST-C component, AI can significantly reduce the variability in IgAN diagnosis and risk stratification. This enhanced objectivity can lead to more consistent patient management, facilitate more robust clinical trials, and ultimately improve outcomes for patients with IgAN. The integration of AI into MEST-C scoring represents a significant step toward precision medicine in IgAN.

DISCOVERING NOVEL HISTOLOGIC PATTERNS AND PROGNOSTIC PREDICTION

Beyond automating the quantification of known histological features, one of the most exciting frontiers for AI in renal pathology lies in its potential to discover novel histological patterns and biomarkers that may not be readily apparent to the human eye. Traditional pathological assessment relies on predefined morphological criteria, but AI, particularly through unsupervised learning and advanced DL architectures, can identify subtle, complex relationships within image data that correlate with disease progression, treatment response, and long-term patient outcomes^[30,31] [Figure 2].

Close

Figure 2:

Clinical impact of artificial intelligence applications in renal pathology.

Export to PPT

UNCOVERING LATENT HISTOLOGIC FEATURES

AI algorithms can process vast amounts of image data from WSIs, identifying intricate textures, cellular arrangements, and spatial relationships that might serve as previously unrecognized prognostic indicators. For example, AI could identify subtle variations in the extracellular matrix composition, the distribution of specific cell types, or the microvascular architecture that are predictive of disease severity or progression, even in cases where conventional histological parameters appear benign or equivocal.^[32-34] This capability moves beyond simply quantifying known lesions to truly discovering new insights into disease pathogenesis.

ENHANCED PROGNOSTIC PREDICTION

The integration of AI-derived histological features with clinical, laboratory, and molecular data holds immense promise for developing highly accurate prognostic models. By combining multimodal data, AI can build comprehensive predictive frameworks that go beyond the capabilities of traditional statistical methods. For instance, AI models can predict the risk of kidney function decline, progression to end-stage renal disease, or allograft failure in transplant recipients with greater precision than current clinical prediction scores.^[35,36] These AI-powered prognostic tools can assist nephrologists in making more informed decisions regarding patient management, including the intensity of follow-up, timing of interventions, and selection of personalized therapies.

IDENTIFYING TREATMENT RESPONSE BIOMARKERS

Another critical application is the identification of histologic patterns that predict response to specific therapies. In many kidney diseases, treatment decisions are empirical, and patients may experience varying responses. AI could analyze biopsy images before and after treatment to identify morphological changes that correlate with therapeutic efficacy, thereby serving as predictive biomarkers for treatment response.^[37,38] This would enable a more precise, biopsy-driven approach to therapy selection, minimizing exposure to ineffective treatments and optimizing patient outcomes. The ability of AI to uncover these novel patterns and integrate them into predictive models represents a significant leap toward truly personalized medicine in nephrology.

CHALLENGES AND FUTURE DIRECTIONS

Despite the immense promise of AI in renal pathology, several challenges must be addressed to facilitate its widespread adoption and integration into routine clinical practice. Overcoming these hurdles will require concerted efforts from pathologists, nephrologists, computer scientists, regulatory bodies, and industry stakeholders [Table 3].

DATA-RELATED CHALLENGES

High-quality, well-annotated datasets are the lifeblood of robust AI model development. Renal pathology datasets often suffer from limitations in both quantity and quality. Variations in tissue processing, staining protocols, and whole-slide imaging platforms across different institutions can introduce technical variability, impacting model generalizability. Furthermore, the meticulous annotation of histological features by expert pathologists, a labor-intensive and time-consuming process, is crucial for training supervised AI models. The scarcity of large, diverse, and uniformly annotated datasets remains a significant bottleneck.^[39-41]

ALGORITHMIC AND TECHNICAL HURDLES

The “black box” nature of some DL models presents a challenge for clinical adoption, as the reasoning behind their predictions can be difficult to interpret. Explainable AI (XAI) techniques are being developed to address this concern by providing insights into model decision-making processes.^[42,43] Ensuring the generalizability of AI models across different patient populations and healthcare systems is another critical concern. Models trained on data from a single institution may not perform as well when applied to data from other centers. In addition, the high computational resources required for training and deploying complex AI models can be a barrier for some institutions.

REGULATORY AND ETHICAL CONSIDERATIONS

Navigating the regulatory landscape for AI-based medical devices is a complex and evolving process. Obtaining approval from bodies such as the Food and Drug Administration (FDA) and Conformité Européenne (European conformity) marking requires rigorous validation and demonstration of clinical utility.^[44,45] Ethical considerations, including patient data privacy, consent, and the potential for algorithmic bias, must also be carefully addressed. Ensuring fairness and equity in the application of AI is paramount to avoid exacerbating existing healthcare disparities.^[46-48] Furthermore, legal frameworks regarding liability for physicians using AI tools are still developing, requiring careful consideration of responsibility and accountability.^[49]

CLINICAL INTEGRATION AND EDUCATION

Integrating AI tools seamlessly into existing clinical workflows is essential for their successful adoption. This requires user-friendly interfaces, interoperability with electronic health records, and clear guidance on how to interpret and act on AI-generated results. Furthermore, there is a need for comprehensive education and training programs to equip pathologists and nephrologists with the skills and knowledge to effectively use and critically evaluate AI-powered tools.^[50,51] Collaborative efforts between academic institutions, professional societies, and industry partners are essential to develop and implement effective training curricula.

CONCLUSION

AI-powered digital pathology stands at the cusp of revolutionizing kidney biopsy interpretation, promising to transform it from a largely subjective art into a more objective, quantitative, and reproducible science. By addressing critical limitations of traditional histopathological assessment, AI offers unprecedented opportunities to enhance diagnostic accuracy, improve prognostic prediction, and facilitate personalized patient management in nephrology. From standardizing sample adequacy and precisely quantifying histological lesions like glomerulosclerosis and IFTA, to improving the consistency of complex scoring systems such as Banff and MEST-C, AI is poised to elevate the quality and utility of renal biopsy reports. Moreover, the ability of AI to uncover subtle, previously unrecognized histologic patterns and integrate multimodal patient data for superior prognostic models represents a paradigm shift towards truly data-driven precision nephrology. While significant challenges remain, particularly concerning data standardization, regulatory pathways, and ethical considerations, the collaborative efforts across pathology, nephrology, and computer science communities are actively addressing these hurdles. The future of renal pathology is undoubtedly digital and AI-augmented, where human expertise is amplified by intelligent algorithms, ultimately leading to better patient outcomes and a deeper understanding of kidney diseases. The journey towards establishing AI-powered digital pathology as the new gold standard for kidney biopsy interpretation is well underway, promising a future where every kidney biopsy yields its maximum diagnostic and prognostic potential.