OpenRounds Editorial
Daily Briefing
Thursday, July 23, 2026
What Changed
A Nature Biomedical Engineering study introduced CLEAR, a radiology foundation model built around clinical concepts with built-in auditability [1]. The architecture matters because it tackles the interpretability gap that has slowed foundation-model adoption in imaging departments.
Research
•[AI Evidence] CLEAR grounds its radiology representations in clinically meaningful concepts rather than opaque embeddings, and the design allows per-prediction auditing of which concepts drove the output [1]. Imaging leaders have seen plenty of foundation models claiming generalizability; this one distinguishes itself by making the model's reasoning inspectable — though real-world validation across sites and modalities is still the open question.
•[AI in Medical Imaging] A multicenter study paired a CNN for liver ultrasound plane recognition with a multimodal LLM that generates textual feedback for the operator, trained on nearly 17,000 images across internal and external validation sets [2]. The hybrid pipeline beat junior and intermediate radiologists in plane identification while producing explanations clinicians can read — a design pattern that points toward ultrasound tools that coach users in real time rather than silently flag errors.
•[AI Evidence] Four leading chatbots answered 20 expert-developed patient-education questions about robot-assisted radical cystectomy, and the evaluation found inconsistent reliability across models on surgical specifics like urinary diversion and complications [3]. Urology programs handing patient education to chatbots should not assume any single model is ready for unsupervised use with complex surgical procedures.
•[AI Evidence] Two LLMs classified open-angle glaucoma from 48 guideline-based cases, with three glaucoma specialists as the reference standard — one model reached higher diagnostic accuracy and classification consistency, but both showed gaps in clinical reasoning quality on the Likert-scored assessment [4]. The study reinforces that diagnostic accuracy alone is a weak proxy for safe clinical deployment; reasoning coherence failed even where answers were right.
Policy & Ops
•[AI in Clinical Operations] Moorfields Eye Hospital is adapting a conversational voice AI called Dora for Turkish-language postoperative cataract follow-up, using focus groups with Turkish-speaking community contributors to shape the design before a forthcoming multilingual trial [5]. The approach — building language access into the AI assistant before piloting rather than retrofitting it — offers a template for systems deploying automated follow-up in multilingual populations without widening disparities.
One to Watch
•[AI in Clinical Operations] A JMIR scoping review synthesizes evaluation frameworks for clinical AI that integrate validation strategies, real-world applicability, and ethical principles into a single structure [6]. AI governance committees looking for a practical scaffold to assess deployed models — beyond accuracy benchmarks — should track whether this synthesis consolidates into an adoptable standard.