Abstract Contrast phases refer to the different stages of blood vessel, tissue, or organ enhancement following the administration of an intravenous contrast agent during CT scans. Contrast phases help distinguish abnormal lesions from normal tissue, where lesions appear hypo- or hyper-vascular, depending on the disease. Therefore, selecting the appropriate phase based on the specific pathology is crucial for accurate detection, meaningful diagnostic interpretation, clinical workflows, and robust quantitative analysis. However, in many polyphasic CT scans, information about phases is missing from the metadata. This study presents a fully automated pipeline to predict the contrast phase of CT scans, using organ-based image features. Our approach has the following novel contributions: (1) Focuses on a limited set of organs (aorta, portal vein, kidney), reducing reliance on extensive organ coverage. This is useful for scans with varying field-of-view where many organs may be absent. (2) Introduces engineered features derived from first-order PyRadiomic statistics, quantifying inter-organ relationships to capture phase-specific intensity patterns. Our pipeline distinguishes non-contrast scans and all three major contrast phases: arterial, portal venous, delayed. Automated segmentation of the aorta, portal vein/splenic vein, and kidneys is performed using TotalSegmentator, a state-of-the-art deep learning algorithm for CT scan segmentation. For each organ mask, a comprehensive set of first-order statistical features, including mean, median, minimum, maximum, standard deviation, skewness, kurtosis, energy, and entropy, is extracted with PyRadiomics to capture intensity distributions. Engineered features reflecting the mean and maximum intensity differences between organs are created, along with the total mean intensity of all three organs, to better distinguish phases. The model was trained and validated on the TOPAZ-1 phase 3 biliary tract cancer trial dataset. TOPAZ-1 comprised of 1418 CT baseline scans (447 portal venous, 396 arterial, 323 non-contrast, 252 delayed) annotated by a radiologist with eight years of experience. For classification, a random forest classifier was developed. Classification achieved an overall accuracy of 0.97, with weighted average precision, recall, and f1-score all at 0.97. Per-phase precision, recall and f1-score all exceed 0.93. On analysis, the engineered features indicate the arterial phase is characterized by a brighter aorta, the venous phase by a brighter portal vein, the delayed phase by increased renal intensity, and non-contrast scans by lower overall intensity relative to the other scans. These patterns align with radiologist expectations and support the model's interpretability. By focusing on organs and features that are most indicative of phase differences, the pipeline yields a simple and interpretable model that achieves accurate polyphasic CT phase identification. Citation Format: Georgia Hughes, Mayank Patwari, Yi Wei, Michael Parker, James Parkin, Zhenning Zhang, Aleksandr Filippov. Automatic contrast phase classification of polyphasic CT scans abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2777.
Hughes et al. (Fri,) studied this question.