Deep Learning Based Detection of High-Risk Renal Masses from All-Cause Abdominal CT

Authors

• Robert Muterspaugh, Diagnostic Radiology Resident - Corewell Health West
• Midhhath Afza Munavar Ali, Diagnostic Radiology Resident - Corewell Health West
• Tairan Song, Michigan State University Computational Mathematics, Science and Engineering Student
• Sabrina L. Noyes, Oncology - Corewell Health West
• Shuyan Zhou
• Conrad Tobert, Urology - Corewell Health West
• Adam Alessio, Michigan State University - Biomedical Engineering
• Emerson Lim, Hematology Oncology - Corewell Health West

Document Type

Conference Proceeding - Restricted Access

Publication Date

5-8-2026

Abstract

Renal masses are commonly detected on CT exams acquired with a variety of contrast timing protocols. Predicting cancer grades of these incidentally detected masses would enable improved outcomes. Numerous prior efforts have applied AI methods for renal prognostics, but these largely rely on optimally timed, kidney CT exams. Hypothesis: Renal mass grading from all-cause CT exams is feasible despite variable contrast timing, and incorporating information from surrounding reference regions will improve prognostic performance compared to main-mass-only features.

This IRB-approved retrospective study curated local CT exams with renal masses resulting in nephrectomy (N=309). CT was acquired with a variety of contrast timing protocols (arterial=24, venous=16, nephrogenic=170, corticomedullary=61, delayed=38). Hand-crafted radiomics features were extracted from mass, normal kidney, aorta, and inferior vena cava (IVC) and used to classify masses as benign (N=44), low-grade (N=176), or high-grade (N=89) cancer. Volumes of interest were segmented with custom nnU-Net (kidney regions) and TotalSegmentator (vascular regions). Multiple feature reduction strategies and classifiers were evaluated. Given the inherent class imbalance in surgically confirmed renal mass cohorts, model performance was using 5-fold cross-validation.

Contrast levels in all-cause CT are more variable than in optimally-timed nephrogenic-phase CT (kidney SD of 35.36 vs 29.40HU; aorta SD of 61.41 vs 23.15HU). The leading classifier (correlation-based feature selection + KPCA+ CatBoost) achieved AUC= 0.62 for benign vs low-grade vs high-grade classification using mass features only. Adding information from the surrounding normal kidney and aorta improved AUC to 0.68, while venous information added no improvement (AUC= 0.67). Similar results were observed for binary classification (benign vs malignant), with AUC of 0.64, 0.74, and 0.71 for mass only, mass+kidney+aorta, and mass+kidney+aorta+IVC.

Contrast timing has an obvious impact on apparent contrast throughout the abdomen. The addition of reference regions improved renal mass grading compared to evaluation of the mass region alone.

Recommended Citation

Muterspaugh R, Munavar Ali MA, Song T, Noyes SL, Zhou S, Tobert C, Alessio A, Lim E. Deep learning based detection of high-risk renal masses from all-cause abdominal CT. Presented at: Research Day Corewell Health West; 2026 May 8; Grand Rapids, MI.

Comments

2026 Research Day Corewell Health West, Grand Rapids, MI, May 8, 2026. Abstract 1968