Radiology ·
AI matches radiology reads for prostate MRI in SCARLET-1; readers and LLMs miss AI-generated chest x-rays; FDA clears MRI-based recurrence tool
Listen to this post
read by AI voices · all episodesTHE SHORT READ
An AI system matched radiology readings for detecting grade group 2 or higher prostate cancer on MRI in diagnostic and screening datasets in the SCARLET-1 study, published Oct. 13 in Radiology.
Other studies found that neither radiologists nor large language models reliably spotted synthetic chest x-rays and that continual learning adapted adult AI radiograph models to neonates. The Food and Drug Administration (FDA) cleared TumorSight Risk, which uses breast MRI to score recurrence risk.
THE DETAILS
- AI Oct 13, 2026In the retrospective SCARLET-1 study published Oct. 13 in Radiology, an AI system was trained and internally tested on 20,471 prostate MRI examinations from 26 cities in 14 countries and externally tested in 1,974 patients. External area under the curve values were 0.86 in the PRIME diagnostic dataset and 0.89 and 0.84 in the STHLM3-MRI and IP1-PROSTAGRAM screening datasets. Agreement with standard-of-care diagnoses, the primary outcome, was 83.0%, 87.9% and 87.7%, meeting prespecified interchangeability criteria in all three. Performance varied substantially across 25 sites, with 28% having an AUC below 0.80.Source: Radiology
- AI Oct 13, 2026A reader study published Oct. 13 in the Journal of Imaging Informatics in Medicine tested radiologists, general practitioners and large language models on 40 real and 120 synthetic chest x-rays. The top performers, Gemini 3.5 Flash at 77.9% balanced accuracy and one radiologist at 75.8%, did not differ significantly (paired difference -2.1 percentage points; 95% CI -5.8 to 1.2). All four human readers detected diffusion-generated images at 82% to 100%, but the radiologists misclassified 38% to 45% of real images, while the more conservative models reached perfect specificity yet detected only 32% to 72% of LLM-generated images.
- AI Oct 13, 2026An international multicenter study published Oct. 13 in Nature Biomedical Engineering repurposed adult AI radiograph models for neonatal care through continual learning, using data from 30 neonatal intensive care units across 20 countries and territories. Continual learning achieved lower anatomical localization error, 0.67 cm, versus 1.29 cm for hospital-specific fine-tuning and 4.30 cm for adult-only models, at 29 of 30 sites. It also improved detection of deep intubations at 27 of 30 sites.Sources: Nature Biomedical Engineering, PubMed
- AI Oct 15, 2026SimBioSys said Oct. 15 that the Food and Drug Administration (FDA) granted 510(k) clearance to TumorSight Risk, which combines breast MRI, pathology and clinical data and is intended as a prognostic indicator of recurrence-free survival. Results can be available before surgery with no additional procedure or tissue shipment, according to AuntMinnie. The software builds on the company's FDA-cleared TumorSight Viz platform.Sources: SimBioSys, AuntMinnie
- Technology Oct 15, 2026A study published Oct. 15 in Clinical Imaging from Memorial Sloan Kettering Cancer Center in New York reviewed 53 biopsies guided by contrast-enhanced mammography (CEM) performed in 2023 and 2024, according to AuntMinnie. Of the biopsied targets, 15 yielded malignancy, four high-risk pathology and 34 benign pathology, and two women had post-biopsy hematomas, none needing intervention. Most targets, 36, were non-mass enhancements, and the researchers found no statistically significant predictors of malignancy.Source: AuntMinnie
The weekly innovation roundup for radiology, published every Sunday for the week before; every item links to its sources. Latest radiology news · All specialties