Evaluate the Full Predicted-Map Training and Inference Pipeline

Evaluate the full semantic-guided multimodal preprocessing pipeline for clear cell renal cell carcinoma grading using predicted nuclei classification maps during both training and evaluation, rather than training on ground-truth maps and applying perturbations only at evaluation.

Background

The proposed method combines RGB histopathology patches with nuclei classification maps before input to a Vision Transformer. The sensitivity analysis evaluates robustness by injecting random segmentation and classification errors into ground-truth maps only at test time, which provides a controlled error sweep but does not reproduce the structured, morphology-dependent errors of an actual upstream nuclei classifier.

The paper explicitly identifies deployment with predicted maps in both phases as unresolved. Such an evaluation is needed to determine whether the reported robustness persists when upstream prediction errors influence both the training distribution and the evaluation inputs, and to assess the performance of the complete practical pipeline.

References

Second, models were trained on ground-truth maps and perturbed only at evaluation, so the reported degradation characterizes robustness to corruption of a reliable input rather than performance with an imperfect model in the loop at both training and evaluation time. Evaluating the full pipeline with predicted maps in both phases is left for future work.

— Semantic-Guided Multimodal Preprocessing for Vision Transformer-Based Clear Cell Renal Cell Carcinoma Grading  (2609.01426 - Javadian et al., 1 Sep 2026) in Section 2, subsection “Sensitivity Analysis”; Section 4, “Conclusion”