Artificial intelligence-enabled medical image analysis in photon-counting CT: A review of segmentation, detection, classification and characterization, and automated quantification

BackgroundPhoton-counting computed tomography (PCCT) provides finely sampled morphological images and energy-resolved products that may alter the behavior of artificial-intelligence models developed for conventional CT.ObjectiveThis review synthesizes downstream AI applications in PCCT and determines which study designs support claims of feasibility, transfer, spatial or spectral contribution, scanner advantage, and clinical utility.MethodsA structured narrative search identified original studies evaluating segmentation, detection, classification and characterization, and automated quantification on reconstructed PCCT images. Each analytical configuration was coded by task, effective input, model provenance, development or adaptation strategy, reference standard, comparator, testing independence, and evidence maturity.ResultsTechnical feasibility has been demonstrated across all four tasks. Established conventional-domain models often retained useful local performance on PCCT, while PCCT-native studies explored virtual monoenergetic images, threshold or energy-bin images, material maps, and multienergy feature spaces. However, small cohorts, repeated non-independent observations, incomplete external testing, limited matched PCCT-EID comparisons, and scarce spatial or spectral ablations restrict causal attribution. Prospective workflow evidence remains uncommon.ConclusionsPCCT creates important opportunities for AI-enabled image analysis, but availability of high-resolution or energy-resolved information does not establish its contribution to performance. Claims should be aligned with the comparator, ablation, reference standard, and validation design required to support them.