30-Second Takeaway
- PET/CT-guided liver ablation with real-time margin and viability checks enables on-table retreatment and excellent local control.
- AI worklist prioritization for primary-care chest radiographs did not shorten time to CT, diagnosis, or treatment for lung cancer.
- Synthetic radiographs already deceive radiologists and LLMs, underscoring the need for deepfake detection and governance.
Week ending March 28, 2026
Oncologic ablation quality, AI limits, deepfakes, and emerging MRI/CT biomarkers
PET/CT-guided colorectal liver metastasis ablation with real-time checks improves local control
This single-center prospective trial tested a PET/CT-guided microwave ablation workflow for 104 colorectal liver metastases in 77 patients. Intraprocedural assessment combined minimal ablation margin measurement, PET/CT for residual uptake, and rapid cytology or viability staining from center and margin biopsies. Fourteen percent of lesions underwent immediate reablation triggered by inadequate margin, residual fluorodeoxyglucose avidity, or viable tumor on rapid biopsy. Margins greater than 5 mm markedly reduced local tumor progression, and no progression occurred when the minimal margin exceeded 10 mm.
AI chest X-ray prioritization does not accelerate lung cancer workup in a large RCT
This multicenter randomized trial evaluated AI worklist prioritization for 93,326 primary-care chest radiographs in the UK lung cancer diagnostic pathway. AI analysis was available in both arms, but AI-based prioritization was toggled on or off by randomization day. Median time to CT was 53 days in both arms, including identical 8-day medians for CTs obtained within 14 days. Lung cancer incidence, time to diagnosis, urgent referral timing, treatment start, and stage at diagnosis did not differ between groups.
Radiologists and multimodal LLMs only moderately accurate at detecting deepfake radiographs
This diagnostic accuracy study compared radiologists and multimodal LLMs in distinguishing GPT-generated synthetic radiographs from authentic images. Seventeen practicing radiologists first reviewed 154 radiographs, including 77 GPT-4o-generated and 77 real images, then repeated classification after being informed about deepfakes. After disclosure, radiologists achieved roughly 70–75% accuracy in distinguishing synthetic from authentic radiographs across GPT-4o and RoentGen datasets. GPT-4o and GPT-5 detected GPT-4o-generated deepfakes more accurately than Llama 4 Maverick and Gemini 2.5 Pro but still missed synthetic images.
References
Numbered in order of appearance. Click any reference to view details.
Additional Reads
Optional additional studies from this edition.