An explainable clinical–radiomics machine learning model for preoperative prediction of WHO/ISUP nuclear grade in clear cell renal cell carcinoma

Comprehensive Summary

The following is a retrospective study that explores explainable machine learning and radiomics models in predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma (ccRCC) using preoperative CT imaging and clinical data. Radiomic analysis extracted 1409 quantitative imaging features from 415 patients with pathologically confirmed ccRCC to train and validate machine learning models. Six major clinical predictors were associated with high-grade ccRCC: BMI, maximum tumor diameter, perirenal filtration, serum albumin, osmolality, and neuron-specific enolase. Researchers tested 12 machine learning algorithms and found Random Forest algorithm to have the strongest overall performance. LASSO feature selection was integrated to filter out redundant variables while retaining key predictors. Radiomic features from CT scans were integrated with LASSO-RF models having the strongest performance (AUC = 0.912, accuracy = 85.3%, sensitivity = 91.7%, specificity = 83.1%). SHapley Additive Explanations analysis improved transparency and interpretability in clinical use by explaining how individual clinical and imaging features influenced model predictions (i.e., NSE or osmolality on probabilities). This combined clinical-radiomic approach suggests that a feasible alternative to accurate but invasive biopsies is in the pipeline for cancer diagnostics.

Outcomes and Implications

Radiomics is a field of medical imaging that uses AI and computational methods to extract quantitative data from CT, MRI, and PET scans. The present study highlights the growing role of explainable AI and radiomics in advancing patient-specific oncology, as seen in this study, where they assessed preoperatively patients with clear cell renal carcinoma. By integrating CT-derived radiomic features with clinical variables, the researchers developed a robust machine learning model for predicting WHO/ISUP tumor grade prior to surgery. Currently, invasive biopsies are required to grade such tumors, but as AI-drive imagining analysis gains traction, risk stratification and personalized surgical consultations can be streamlined and efficient. Clinicians can potentially leverage an AI-assisted diagnostic infrastructure to save time, enhance grading accuracy, and ultimately improve outcomes in oncological patients. For healthcare systems, AI infrastructure can unlock efficiencies not previously possible without AI-radiomic systems. While this study has a relatively limited sample size, a lack of standardized radiomic workflow across institutions, and is a retrospective study, it nonetheless proves that there is high potential for noninvasive cancer grading through radiomics and AI-powered means.