Optimizing renal mass histology prediction with artificial intelligence using handcrafted radiomics.

purposeTo determine the histology of a renal mass, physicians cannot rely on imaging alone. Radiomics—the extraction of quantitative data from medical imaging—could help address this challenge. This study aimed to build a classification model using CT-scan radiomics and/or clinical features to distinguish clear cell renal cell carcinoma (ccRCC) from other histologies.

methodsWe included patients who underwent surgery for suspected localized RCC. A total of 345 masses were included (72% ccRCC, 28% other histologies). Clinical data were extracted and each renal mass was manually segmented on CT and CECT (contrast enhanced CT). Then, 171 radiomics features were extracted from each lesion. The dataset was randomly split into a learning set (80%) and a hold-out set (20%). The learning set was further portioned into 10 subsets; each subset was split into training (80%) and testing (20%) sets and used to train and test 10 independent machine learning models using the XGBoost algorithm. The best model was subsequently validated on the hold-out set.

resultsThe CECT-based radiomics model showed the best performance: AUC = 0.80, sensitivity = 67% and specificity = 77%. Adding clinical features to the model did not improve its performance. Using the hold-out set, the final model achieved an AUC of 0.88, with 56% sensitivity and 93% specificity.

conclusionsOur radiomics-based classification model showed high performance in differentiating ccRCC from other histologies using CECT scans. These findings support the role of radiomics as a potential non-invasive tool in kidney cancer management. However, external validation on independent cohorts is needed before clinical application.