30-Second Takeaway
- Photon-counting CT delivers cross-domain image and dose gains, with emerging evidence for combined anatomic–functional lung assessment.
- AI for pulmonary nodules and coronary plaque can boost accuracy and prognostication, but model training and validation remain fragile.
- Propensity-adjusted data and synthetic contrast studies both argue for cautious, selective use of iodinated and gadolinium agents.
- Radiomics models are often undertrained, so “externally validated” tools may be less reliable than their performance metrics suggest.
- LLM explanation format and standardized hybrid infection imaging workflows materially influence diagnostic accuracy and communication.
Week ending April 25, 2026
Photon-counting CT and AI tools are maturing, but contrast risk and underpowered models still constrain routine adoption
Photon-counting CT shows broad clinical gains but needs standardized protocols
This Radiology review summarizes more than 450 photon-counting CT (PCCT) papers across cardiac, thoracic, neurovascular, musculoskeletal, abdominal, and pediatric imaging. Core PCCT features—better spatial and contrast resolution, spectral imaging, and lower electronic noise—translate into improved image quality and dose efficiency. Reported benefits include higher diagnostic performance and early signals for impact on disease management and cost-effectiveness. The review outlines protocol optimization strategies and workflow issues that currently limit scalable implementation. It stresses the need for protocol standardization and multicenter trials before widespread replacement of energy-integrating CT systems.
DeepFAN AI assistance improves junior radiologists’ pulmonary nodule classification
DeepFAN is a transformer-based model trained on over 10,000 pathology-confirmed pulmonary nodules for benign–malignant classification. In a multicenter trial of 400 CT cases, DeepFAN achieved AUCs of 0.939 on an internal test set and 0.954 on the clinical dataset. Assisting 12 junior radiologists, DeepFAN increased average AUC by 10.9% and accuracy by 10.0%, with parallel gains in sensitivity and specificity. Interreader agreement improved from fair to moderate, suggesting AI support may homogenize nodule workup and reduce unnecessary follow-up.
PlaqueSegNet automates CCTA plaque quantification and predicts MACE risk
PlaqueSegNet is a fully automated deep learning model for coronary plaque volume quantification on CCTA, trained on 1409 patients from 17 hospitals. Across four external datasets, including IVUS comparisons and photon-counting CT, plaque volumes showed excellent agreement with experts, with intraclass correlation coefficients above 0.90. PlaqueSegNet-derived plaque volumes predicted major adverse cardiac events with C-indices around 0.64–0.65 in two CCTA cohorts and 0.74 in a serial CCTA cohort. These results support integrating automated plaque quantification into CCTA reporting to add prognostic information beyond stenosis severity.
Critical review finds meaningful AKI and dialysis risk after CECT in advanced CKD
This review reappraises propensity-score-adjusted studies assessing contrast-enhanced CT (CECT)–associated kidney injury. Among emergency and inpatients with chronic kidney disease, CECT was linked to higher AKI rates in most cohorts, with a weighted absolute risk increase of 4 percentage points at GFR <30 mL/min/1.73 m2. Dialysis risk also increased in this group, with a weighted absolute risk increase of 1.8 percentage points at GFR <30 mL/min/1.73 m2. The authors note selection bias and masking by acute kidney recovery may explain several “no-risk” reports and caution against dismissing CECT nephrotoxicity in high-risk CKD.
References
Numbered in order of appearance. Click any reference to view details.
Additional Reads
Optional additional studies from this edition.