Morphological analysis, including the determination of renal cell carcinoma (RCC) histotype, tumor grade, presence of lymphovascular invasion, tumor necrosis, and sarcomatoid dedifferentiation, is pivotal for RCC diagnosis. It not only defines prognosis but also predicts the impact of eventual systemic treatments. In contemporary practice, this analysis must be complemented by genetic and cytogenetic assessments.¹ However, RCC histological diagnosis and classification can pose challenges due to its encompassment of a diverse range of histopathological entities, which have recently undergone revisions.²

Over the years, the diagnostics in RCC have evolved through the integration of modern counterparts such as electronic health records, digitalized radiology, and virtual pathology. This evolution has generated a huge amount of data, which can be processed using characterization algorithms or artificial intelligence (AI).³

The use of AI in RCC histopathology, known as pathomics or computational pathology, is relatively new. AI can assist the pathologist in improving efficiency, accessibility, cost-effectiveness, and time consumption. It also enhances accuracy and reproducibility, reducing subjectivity. Additionally, whole slide imaging technology, which refers to scanning of conventional glass slides in order to produce digital slides, enables machine learning in pathology by providing a vast amount of high-quality information for training and testing AI models to identify specific features and patterns that may be challenging for the human eye to discern.⁴

Machine learning, a subfield of AI, utilizes algorithms that enable computers to learn from digital images of tissue samples. In histopathology, it can be employed for the digital analysis of images to identify cell types, different structures, and to segment various regions of a given tissue sample.⁵ The capabilities of machine learning have advanced with the development of deep learning: a section of machine learning is now focused on creating virtual neural networks, drawing inspiration from the ways in which the neurons of a human being communicate.⁵ Deep learning models are adept at extracting features and learning from data. They can automatically identify complex patterns and relationships within diverse, large datasets, such as those used in cancer diagnostics.

However, choosing the best algorithm for the application of AI in histopathology is challenging. There are 3 primary types of learning: (1) supervised learning, which utilizes labeled data for training; (2) unsupervised learning, which identifies patterns without labels; and (3) weakly supervised learning, which strikes a middle ground by using partially labeled data.⁶

In our daily routines, we are well aware of repetitive and time-consuming tasks, such as the analysis of high-volume biopsy tissue samples and the counting of lymph nodes yielded during surgery. In such cases, AI has the potential to flag suspicious regions for inspection and might enable autonomous assessment. Additionally, AI can assist the pathologist in tasks like classifying different regions of cancer based on varying tumor grades using color-coded highlighting.

Moreover, by combining segmentation, detection, and classification techniques, it becomes possible to objectively quantify established biomarkers used in clinical practice. Notably, in the field of RCC pathology, specific instances include the evaluation of tumor-infiltrating lymphocytes and the quantification of PD-L1 (programmed death-ligand 1)–positive cells, which can even be predicted directly from slides.^7,8 Therefore, AI might aid in a wide spectrum of tasks ranging from tumor detection and classification to predictive and prognostic modeling.

Where are we now? Some authors have developed deep learning–based algorithms for RCC diagnosis, subtyping, and grading on biopsy specimens. However, they primarily focused on identifying the main subtypes of RCC without considering benign tumors.⁹ Using specimens obtained from surgical resection, other authors have achieved promising results by employing AI in differentiating among RCC subtypes and normal parenchyma.⁶ Undoubtedly, the pioneer experiences are witnessing how AI and machine learning in RCC pathology hold promise for the future of RCC diagnosis. They might help us overcome several issues faced by the pathologist with “traditional” histopathology, primarily concerning time consumption and intra-/interobserver variability.

A representation of the ideal pathway we imagine for the development of pathomics algorithms is summarized in the Figure: following either a biopsy or surgical resection, a whole slide image is created and derived patches are used through a digital scanner to train the algorithm in defining diagnostic and prognostic models.

One issue could lie in the fact that supervised learning-based algorithms could lead to the so-called “black box”: while these algorithms are efficient at performing the assigned tasks, the generated outputs cannot be visually authenticated (ie, a human cannot oversee them; thus, the pathologist must have faith in the findings). Up to date, the available AI algorithms are either noninferior or even outperform the standard methods, but it is important to note that most technologies are currently unavailable for widespread clinical use, and further evidence is needed.

In our opinion, in the immediate future, AI will assist the community of uropathologists in elevating the average quality of assessments, particularly in recognizing different tumor gradings rather than different histotypes.

We believe this is the major need. From our clinical experience as “second-opinionists,” we have observed that the majority of misclassifications with clinical implication occur in assigning tumor grade.

In a more distant future, the perspective of pathomics could lie in aiding the prediction of RCC prognosis. This will be particularly significant for the uro-oncologist, as an “augmented intelligence,” relying on extensive big data (including tumor molecular characteristics), could differentiate between apparently similar pT1b grade 3 clear cell RCCs that will have different natural histories.