Renal Cell Carcinoma subtyping: Learning from multi-resolution localization

Renal Cell Carcinoma (RCC) is often diagnosed at advanced stages, limiting treatment options. Since prognosis depends on tumour subtype, accurate and efficient classification is essential. Artificial intelligence tools can assist diagnosis, yet their dependence on large annotated datasets hinders broader adoption. This study investigates a Self-Supervised Learning (SSL) framework that exploits the multi-resolution structure of Whole histological Slide Images (WSIs) to reduce annotation requirements while maintaining reliable diagnostic performance.

We developed a SSL model inspired by the pathologist's multi-scale reasoning, integrating information across magnification levels. Robustness and generalization were evaluated through an external validation on a public RCC benchmark and one internal validation using cohorts from the same institution but collected in different periods, with distinct scanners and laboratory workflows.

The proposed SSL approach demonstrated stable classification performance across all validation settings, reducing dependence on manual labels and improving robustness under heterogeneous acquisition conditions. These findings support its potential as a generalizable and annotation-efficient strategy for RCC subtype classification.