Granular Machine Learning-Based Computed Tomography Contrast Phase Prediction

To develop and evaluate a machine learning framework that detects intravenous contrast and distinguishes eight granular renal contrast phases on abdominal computed tomography (CT) scans to improve renal assessment.

This retrospective study included abdominal CT scans obtained at Mayo Clinic from January 1, 2001, to December 31, 2009. In total, 3033 scans from 1017 patients with renal cell carcinoma were included. A ConvNeXt-Femto deep learning (DL) model with dual output heads was trained for contrast detection and renal contrast phase prediction using binary classification and regression objectives, respectively. A random forest (RF) regression model was trained on DL-extracted features to predict 8 fine-grained phases spanning early to late corticomedullary, nephrographic, and pyelographic. Model performance was further evaluated using an internal-external cohort of abdominal CT scans from January 1, 2010, to December 31, 2020, comprising of 8856 series from 4760 patients.

The DL classifier detected contrast presence with 100% accuracy. The DL-only regression model reached a mean absolute error of 0.34, compared with 0.29 for the hybrid DL+RF model. Agreement analysis between the models' ensemble and 2 radiologists reported reliability, with κ values of 0.86 for predicting the exact category, 1.00 for neighboring categories, and 0.98 for super-category grouping. Internal-external validation indicated that the model successfully operated across datasets differing in patient cohort and imaging characteristics.

This DL+RF framework enables automated granular renal contrast phase discrimination and reduces inter-rater variability, representing a meaningful advancement in artificial intelligence-assisted abdominal CT interpretation and supporting improved patient care.