Kidney Histopathology Eval

kidney-histopathology-eval

A Comprehensive Benchmark of Histopathology Foundation Models for Kidney Digital Pathology Images — Kasireddy et al. (2026) (arXiv:2603.15967, 2026)

What this evaluates

Evaluates histopathology foundation models on kidney-specific downstream tasks, including tile-level morphological classification, molecular information estimation, and slide-level diagnostic/prognostic inference across diverse staining protocols (H&E, PAS, PASM, IHC).

Datasets

• Kidney Digital Pathology Benchmark — total ?; splits: test (-1)

Metrics

• Matthews Correlation Coefficient (MCC) (primary) — range: [-1, 1]
  • Measures the quality of binary classifications by accounting for true/false positives and negatives. Formula: (TPTN - FPFN) / sqrt((TP+FP)(TP+FN)(TN+FP)(TN+FN)). Ranges from -1 to 1.

Input / output format

Input: Frozen embeddings from Histopathology Foundation Models (HFMs) derived from tile-level or slide-level kidney histopathology images, paired with ground-truth labels for tasks such as glomerulosclerosis classification, GBM spike detection, tubule classification, and inflammation classification.

Output: Predicted class labels or regression values generated by linear probing or k-Nearest Neighbors (kNN) classifiers applied to the frozen HFM embeddings.

Scoring recipe

def compute_mcc(y_true, y_pred):
    tp = sum((y_true == 1) & (y_pred == 1))
    tn = sum((y_true == 0) & (y_pred == 0))
    fp = sum((y_true == 0) & (y_pred == 1))
    fn = sum((y_true == 1) & (y_pred == 0))
    numerator = (tp * tn) - (fp * fn)
    denominator = math.sqrt((tp + fp) * (tp + fn) * (tn + fp) * (tn + fn))
    return numerator / denominator if denominator > 0 else 0.0
# Aggregate across 1000 bootstraps of test-time predictions from repeated stratified group 5-fold CV.

Common pitfalls

• Models are evaluated using both linear probes and k-NN classifiers, which can yield different performance rankings and should not be conflated.
• Performance is reported as bootstrap statistics (median/IQR) rather than single-point estimates, requiring careful interpretation of confidence intervals.
• Tile-level tasks use stratified group k-fold cross-validation to prevent data leakage from the same patient or slide across folds.

Evidence (verbatim from paper)

Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated stratified group five-fold cross-validation runs.

Citation

@misc{kasireddy2026kidneyhistopathologybenchmark,
  title={A Comprehensive Benchmark of Histopathology Foundation Models for Kidney Digital Pathology Images},
  author={Kasireddy et al. (2026)},
  year={2026},
  note={arXiv:2603.15967}
}

• arXiv: 2603.15967