Answers are grounded in this paper and cite their sources.
Background Clear cell renal cell carcinoma exhibits substantial molecular heterogeneity, and clinically relevant alterations involving BAP1, SETD2, the PI3K/AKT/mTOR signaling pathway, and hypoxia-related protein-expression profiles may not be fully captured by conventional imaging interpretation. Radiogenomic neural networks may support non-invasive molecular risk stratification; however, limited explainability, small validation cohorts, and uncertain transportability remain important obstacles. Objective This study developed an explainable radiogenomic neural network for molecular high-risk stratification in clear cell renal cell carcinoma and secondarily evaluated a separate breast cancer cohort only as a descriptive workflow stress test. Methods This retrospective single-center exploratory study enrolled 210 patients treated between January 2019 and December 2025. The primary renal cohort comprised 150 patients with clear cell renal cell carcinoma and was chronologically divided into a development set, an internal validation set, and a small temporal test set. A separate cohort of 60 patients with invasive breast cancer was retained only as a secondary descriptive workflow stress test. Because the breast image-risk score generation workflow could not be reliably reconstructed, the breast analysis was not treated as external validation or a reproducible renal-to-breast model-transfer experiment. The renal molecular high-risk endpoint was a prespecified composite of BAP1 protein loss, high P4HA3 expression using a development-set-fixed cutoff, pathogenic or likely pathogenic alterations in BAP1 or SETD2, or pathogenic/likely pathogenic alterations involving the mTOR signaling pathway. Endpoint-defining variables were excluded from predictors. A three-dimensional convolutional neural network generated the image risk score and represents the fully documented primary predictive model in the present revision. Archived combined risk scores that integrated the image risk score with prespecified non-endpoint variables were retained only as a secondary exploratory analysis because the estimator-level fusion operator and associated scaling/encoding details could not be completely reconstructed from the archived materials. Grad-CAM and SHAP were used as post hoc explanation methods. Results Molecular high-risk status was identified in 43 of 150 renal patients (28.7%), including 32/105 (30.5%) in the development set, 8/30 (26.7%) in internal validation, and 3/15 (20.0%) in the temporal test set. In renal internal validation, the fully documented image-only 3D CNN achieved an AUC of 0.756; the secondary archived combined-model AUC was 0.801. In the renal temporal test set, the image-only AUC was 0.750 and the secondary combined-model AUC was 0.639; because this set contained only 15 patients and 3 high-risk events, these estimates were considered exploratory. Threshold-based classification metrics and calibration statistics for the primary image-only 3D CNN could not be reliably reconstructed from the archived case-level materials and therefore are not reported. In the breast workflow stress-test cohort, the image-risk AUC was 0.652 and the archived combined-score AUC was 0.649. Because the breast image-risk score generation workflow could not be reliably reconstructed, breast performance was retained only as descriptive technical stress-test output and not as evidence of a reproducible renal-to-breast model-transfer experiment. Conclusion An explainable radiogenomic 3D CNN showed measurable discrimination for molecular high-risk status in clear cell renal cell carcinoma during internal validation. The image-only model is the principal reproducibly documented predictive model in this revision, whereas the archived combined-model outputs are retained as secondary exploratory analyses because the exact estimator-level fusion implementation could not be fully reconstructed. Numerically different performance in the 15-patient temporal test set should be regarded as exploratory because only three high-risk events were present, and the estimates may be sensitive to the single development split and initialization. The breast analysis was retained only as a descriptive technical stress test because its image-risk score generation workflow could not be reliably reconstructed; it was not considered external validation of renal cancer biology or a reproducible model-transfer experiment. Larger multicenter cohorts with repeated model refitting, complete prospective archiving of model specifications, and prospective testing are required before clinical translation.
No takes yet. Share an insight, caveat, or question.
Jia et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: