RenalPath

The 2022 WHO classification recognizes 25 subtypes of renal cell tumor, but the WHO/ISUP nuclear grade carrying most of the prognostic weight in a renal pathology report is validated for only two — clear cell and papillary RCC. A patient with a chromophobe or TFE3-rearranged tumor leaves the microscope with a subtype and no graded risk, and where grading does apply it is unreliable: a prior study put the interobserver kappa between two pathologists at 0.45.

RenalPath answers this with a cascade of four supervised heads over a frozen tile encoder. Slides are normalized to roughly 1 µm per pixel, cut into non-overlapping 224 × 224 patches, and embedded by Prov-GigaPath into 1536-dimensional vectors that are never updated. The first head classifies each patch into five tissue components — normal, tumor, pseudocapsule, necrosis, and sarcomatoid differentiation — and its output doubles as a region filter: normal patches are discarded and the remaining four classes pooled into a tumor-associated macro-region. Subtype classification, nuclear grading, and the whole-slide risk score all read that curated region rather than the raw slide, so one annotation effort feeds every downstream task — including risk stratification for the subtypes no grading system covers.

The system was developed by the Department of Urology at Zhongshan Hospital, Fudan University with Fudan's Digital Medical Research Center and multi-center collaborators, and published in BMC Medicine in August 2026. Its code and checkpoints ship under the name full-stackRCC; RenalPath is the name under which the group registered it as the index test of a prospective trial (ChiCTR2600116157). It is unrelated to RenalPath, a self-supervised kidney-biopsy encoder from a separate group that shares the name.

#Key Features

• Four clinical tasks from one frozen backbone: tissue-region detection, nine-way subtype classification, binary WHO/ISUP grading, and prognostic risk scoring are separate heads over the same patch embeddings, so a slide is featurized once.
• Tissue detection as a region filter: the five-class patch head is both a deliverable — necrosis and sarcomatoid differentiation are reportable prognostic features — and the step that defines the tumor-associated region for the other three tasks.
• Prognosis beyond the graded subtypes: the WSI-based Risk Score is fit across all renal cell tumor subtypes, reaching tumors WHO/ISUP grading does not cover.
• Thresholds frozen before external testing: for patch-level and slide-level classification, decision thresholds were chosen by maximizing Youden's J on the training set, then applied unchanged to the validation and test cohorts; the remaining tasks use a fixed 0.5 probability threshold in every cohort. All four trained checkpoints are published, so the system runs without refitting.

#Technical Details

Patch embeddings come from Prov-GigaPath, a ViT pretrained with DINOv2 on 1.3 billion Providence-network patches — a corpus containing no TCGA or CPTAC data, ruling out leakage into the public validation cohorts. The two patch-level heads are three-layer MLPs with batch normalization and ReLU, trained with SGD and cross-entropy; the grading and prognosis heads use attention-based multiple instance learning over per-patient bags, trained with AdamW. The study drew on 11,135 WSIs from 7,033 patients across Zhongshan, Xiamen, Zhangye and Huadong hospitals plus TCGA-KIRC, TCGA-KIRP, TCGA-KICH and CPTAC-CCRCC; Huadong was held out entirely as the external test cohort (896 slides, 877 patients), leaving 8,154 training and 2,085 internal validation slides. Tissue detection reached AUCs of 0.967–0.994 on internal validation, falling to 0.922 for sarcomatoid differentiation externally, and slide-level subtype macro-AUC was 0.983 internally against 0.928 externally. Binary grading reached AUC 0.867 (0.911 for ccRCC, 0.776 for pRCC), agreeing with subspecialist consensus on 86.4% of cases against 81.4% for general pathologists. The WRS beat WHO/ISUP grade on recurrence-free survival (C-index 0.735 vs 0.694 externally) and disease-specific survival (0.743 vs 0.721), while remaining below TNM stage, which encodes tumor size and fat invasion no slide reveals.

#Applications

The immediate use is decision support inside a renal tumor reporting workflow: pre-marking necrosis and sarcomatoid regions that must be recorded, proposing a subtype and flagging out-of-distribution cases that warrant immunohistochemistry, FISH or sequencing, offering a second read on nuclear grade, and attaching a risk score to subtypes that currently receive none. Since it consumes only H&E slides already scanned for routine diagnosis, it also suits annotating archival collections for outcomes research.

#Impact

RenalPath shows that a general pathology encoder plus modest task heads can cover an entire organ-specific reporting pipeline rather than a single endpoint, and its pan-subtype risk score addresses a gap grading has left open since the subtype count grew. The caveats are real. Feature extraction requires Prov-GigaPath, whose weights are access-gated, and the released inference script hardcodes cluster paths that must be edited before it runs elsewhere; the patch-tiling preprocess.py the README points to is not in the repository. The training slides are not public, the grading head was fit on a curated 1,000-patient consensus cohort, and external performance drops for rarer categories — papillary RCC sensitivity falls to 0.649 at Huadong. The authors state plainly that prospective workflow studies are needed; the registered trial comparing the model against PLIP, UNI and senior pathologists has not reported.