Optimizing renal mass histology prediction with artificial intelligence using handcrafted radiomics

Date

2026-02-25

Authors

• Teodora Boblea Podasca¹
• Mahdi Loutfi²
• Marc-Antoine Blais³
• Marie-Lou Gadoury-Campbell³
• Stéphanie Boulet⁴
• Martin Vallières²
• Patrick O. Richard¹

¹ Division of Urology, Department of Surgery, Centre Hospitalier Universitaire de Sherbrooke, Sherbrooke, Canada

² Department of Computer Science, Université de Sherbrooke, Sherbrooke, Canada

³ Faculty of Medicine and Health Sciences, Université de Sherbrooke, Sherbrooke, Canada

⁴ Division of Urology, Department of Surgery, Centre Hospitalier de l’Université Laval, Québec, Canada

Abstract

Purpose

To 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.

Methods

The study included patients who underwent surgery for suspected localized renal cell carcinoma. A total of 345 masses were included, 72% of which were ccRCC and 28% of which had other histologies. Clinical data were extracted, and each renal mass was manually segmented on CT and contrast-enhanced CT (CECT). A total of 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 divided into 10 subsets, each 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.

Results

The CECT-based radiomics model showed the best performance, with an area under the curve (AUC) of 0.80, sensitivity of 67%, and specificity of 77%. Adding clinical features to the model did not improve its performance. On the hold-out set, the final model achieved an AUC of 0.88, with 56% sensitivity and 93% specificity.

Conclusions

The 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.

Links

• World Journal of Urology