Granular Machine Learning-Based Computed Tomography Contrast Phase Prediction.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Radiological Assessment
• Computed Tomography (CT) Image Analysis

Background:

• Accurate assessment of renal contrast phases in abdominal CT is crucial for patient care.
• Manual interpretation of renal contrast phases can be time-consuming and subject to inter-rater variability.
• Existing methods may lack the granularity needed for precise renal assessment.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) framework for detecting intravenous contrast.
• To distinguish eight granular renal contrast phases on abdominal CT scans.
• To improve the accuracy and consistency of renal assessment in CT imaging.

Main Methods:

• Retrospective study using 3033 abdominal CT scans from 1017 patients with renal cell carcinoma.
• A ConvNeXt-Femto deep learning (DL) model was trained for contrast detection and phase prediction.
• A hybrid DL+Random Forest (RF) model utilized DL-extracted features for fine-grained phase prediction (8 phases).

Main Results:

• The DL classifier achieved 100% accuracy in contrast detection.
• The hybrid DL+RF model demonstrated a mean absolute error of 0.29 for phase prediction, outperforming DL-only regression (0.34).
• High agreement (κ values 0.86-1.00) was observed between the model ensemble and radiologists, with successful internal-external validation.

Conclusions:

• The developed DL+RF framework enables automated, granular discrimination of renal contrast phases.
• This AI-assisted approach significantly reduces inter-rater variability in abdominal CT interpretation.
• The framework represents a meaningful advancement supporting improved patient care through enhanced CT analysis.