Weekly AI digest

Three things that mattered

• LLMs move into clinical radiology reporting

  MIRA, fine-tuned on 1.87 million reports from 42 hospitals, outperformed GPT-4o for generating radiology impressions, judged at least equivalent to references in 69.0% of blinded comparisons by 24 radiologists. DeepSeek-R1 reached 84.4% error detection in emergency reports, comparable to senior radiologists.

• Two regulatory and acquisition moves reshape the market

  Philips obtained FDA 510(k) clearance for the Spectral CT Verida on 16 April, while Sectra finalized its acquisition of Oxipit, maker of the CE class IIb autonomous chest X-ray tool ChestLink, on 15 April.

• The FDA holds a strict line on AI device review

  The FDA rejected Harrison.ai's petition to exempt six categories of radiology AI devices (CADe, CADx, CADt and combined) from the 510(k) process, confirming Predetermined Change Control Plans as the recommended path. In 2025 it granted at least 255 such clearances.

Peer-reviewed

• MIRA: fine-tuned language model for automated radiology impression generation, multicenter evaluation

  Radiology: Artificial Intelligence (RSNA) · First Hospital of Jilin University, China, 15 April 2026

• Error detection in emergency radiology reports by large language model: a multistep evaluation study

  Journal of Medical Internet Research · First Affiliated Hospital, Jinan University, Guangzhou, 14 April 2026

• GPT-4o for automated determination of follow-up examinations from routine clinical radiology reports

  Scientific Reports · University of Cologne and University Medical Center Mainz, 16 April 2026

• Diagnostic accuracy of an AI tool for appendicular skeleton fracture detection on radiographs, an independent bone-level study

  European Radiology · Radiological Artificial Intelligence Testcenter (RAITC), Copenhagen, 17 April 2026

• Mammography-based AI model for predicting axillary nodal status after neoadjuvant therapy in breast cancer

  European Radiology · Kunming Medical University and Shanghai Jiao Tong University, 18 April 2026

• Automated longitudinal tracking of pancreatic cysts from radiology reports by hybrid LLM and rules algorithm

  Abdominal Radiology · NYU Langone Health and Columbia University Irving Medical Center, 17 April 2026

• Artificial intelligence in hepatocellular carcinoma: imaging subtyping, treatment-response prediction and prognosis

  Abdominal Radiology · Memorial Sloan Kettering Cancer Center, New York, 16 April 2026

Industry & regulation

• Philips obtains FDA 510(k) clearance for the Spectral CT Verida

  16 April 2026 (K253649, decision 27 March 2026). First detector-based spectral CT fully driven by AI, with always-on dual-layer spectral detection and AI reconstruction at 145 images per second. Already CE marked since November 2025.

• Sectra finalizes its acquisition of Oxipit

  15 April 2026. Oxipit is a pioneer of autonomous radiology AI and maker of ChestLink, the first CE class IIb solution for autonomous analysis of chest radiographs, which automatically sets aside high-confidence normal studies.

• HeartFlow files a patent infringement suit against Cleerly

  13 April 2026 (case 2:26-cv-00292, Eastern District of Texas). HeartFlow alleges Cleerly products (Ischemia, Plaque Analysis, Compare) infringe six patents on AI with CT imaging for non-invasive coronary diagnosis. Cleerly founder James Min consulted for HeartFlow from 2012 to 2017.

• FDA rejects Harrison.ai's 510(k) exemption petition for radiology AI

  Formal decision dated 1 April 2026, published this week. The FDA declined to exempt six categories of radiology AI (CADe, CADx, CADt and combined) from 510(k), receiving 47 comments, and confirmed Predetermined Change Control Plans as the recommended path. At least 255 such clearances were granted in 2025.

• ACR tracks the wave of state AI legislation in medical imaging

  The ACR is monitoring a series of state-level bills on radiology AI covering algorithmic transparency, bias, AI-related liability and clinical oversight, while developing national performance-tracking infrastructure via its Assess-AI program and the ARCH-AI platform.

• SIIM-ACR Data Science Summit 2026 convenes on AI governance

  The summit brings together imaging informatics and radiology decision-makers around AI governance, performance monitoring, algorithmic transparency and responsible implementation. Available virtually and in person, with up to 5 AMA PRA Category 1 credits.

Conferences & voices

Sectra chief executive Jan Kronander framed the Oxipit acquisition as adding autonomous AI that can genuinely reduce radiologists' workload in daily practice, together with the clinical validation built over years of real-world testing. NYU Langone radiologists Chenchan Huang and Yiqiu Shen, reporting in Abdominal Radiology, showed that a hybrid LLM and rule-based algorithm matched dominant pancreatic cysts across serial reports with 96% agreement, while cyst growth classification varied with the guideline framework used, underscoring how algorithm design choices shape real-world surveillance. Media coverage from STAT News and MedCity News emphasized the gap between improving detection and lagging clinical follow-up, including the financing barrier to reporting opportunistic coronary calcium on the roughly 19 million lung CT scans performed annually in the United States.

Source articles are indexed in PubMed with verified DOIs. Manuscript-stage work is excluded. The full weekly report is produced in French.

Get the weekly digest

One email a week on what matters in radiology and imaging AI. Researched by a practicing radiologist. Free, no vendor sponsorship.

By subscribing you agree to receive the weekly digest by email. Unsubscribe anytime. No sharing with third parties.