A newsletter on the latest in AI for healthcare.
Welcome back,
In this issue, RadAgent makes chest CT interpretation more traceable, while today’s FDA-cleared prostate MRI AI story shows specialised imaging tools moving deeper into clinical workflows.
The model story is Hygieia, a rare-disease AI agent for diagnosis and gene prioritisation. On the commercial side, Blank Bio, Optura, and Nexalin show AI moving further into oncology, health-plan operations, and digital neurocare.
Here is what you need to know.
SUMMARY
Top Research Paper
A reinforcement-learning-trained radiology agent turns chest CT report generation into a stepwise, tool-grounded process that raises accuracy and creates a more inspectable audit trail.
Top AI News
DeepHealth’s Prostate AI has received FDA 510(k) clearance and CE Marking, extending the regulator's steady cadence of specialised imaging clearances.
Top Model
Hygieia is a router-based, multimodal AI agent for rare-disease diagnosis and risk-gene prioritisation, with confidence scoring and reasoning traces.
Bedside Bets
Startup rounds, deals, and moves.
Blank Bio raises $7.2M to build RNA foundation models for precision oncology: Built as an applied AI research lab, Blank Bio uses foundation models to learn from tumour RNA data and support biomarker discovery, trial design and clinical diagnostics.
Optura raises a $17.5M Series A, bringing total funding to over $25M: Its enterprise healthcare AI platform helps health plans and providers prioritise AI use cases, turn workflows into AI agents and track return on AI.
Nexalin closes a $1.3M stock-based acquisition of PONM: Best known for non-invasive neurostimulation, Nexalin is adding PONM's AI-integrated remote monitoring, clinical data capture and virtual-care infrastructure to its neurocare platform.
Pulse Check
Other quick reads across health AI.
Cervi-LLM uses multimodal AI to locate cervical lesions, stratify disease severity and guide real-time biopsy during colposcopy.
The Department of Health – Abu Dhabi and J&J are building an intelligent operating room network that captures surgical video and multimodal data to support case review and train surgical AI.
TOP PAPER
🩻 RadAgent makes chest CT reports easier to check
Source: arXiv · 16 April 2026 · ETH Zurich, University of Zurich, Stanford AIMI and collaborators
Most 3D vision-language models for CT reporting give clinicians a finished paragraph with little visibility into how the model reached its conclusion. RadAgent takes a different route. It wraps the CT-Chat VLM in a clinician-style diagnostic checklist and ten specialised tools, so each intermediate finding can be inspected, validated, or challenged.
That makes the work especially relevant for radiology teams trying to move from black-box reporting toward AI systems with clearer reasoning trails.
Question
Can RadAgent move clinicians from passive readers of final CT reports to active reviewers of the model's reasoning, while making chest CT reporting more accurate and trustworthy?
Source Roschewitz et al
Approach
RadAgent uses an instruction-tuned Qwen3-14B as the agent policy, fine-tuned with LoRA and GRPO to run a ReAct loop over a nine-item diagnostic checklist.
Its main tools include CT-Chat for draft report generation and 3D CT VQA, google/gemma-3-27b-it for 2D slice-level VQA, CT-CLIP VocabFine for 18-pathology classification, and TotalSegmentator for anatomy and effusion segmentation.
Additional tools handle slice selection and CT windowing, while Qwen3-30B-A3B-Thinking helps score abnormal findings during reward design.
Trained on CT-RATE, evaluated on CT-RATE validation and test splits and on the external RadChestCT set; robustness and faithfulness were measured through a 1,000-case hint-injection experiment.
Results
On CT-RATE test: +6.0 macro-F1 points and +5.4 micro-F1 points over CT-Chat.
Strongest improvements came on previously low-performing pathologies, including pleural effusion, consolidation, cardiomegaly, and lymphadenopathy.
Robustness to incorrect hint injection rose to 83.7% versus 58.9% for CT-Chat.
Faithfulness reached 37.0% versus 0.0% for CT-Chat, which never acknowledged hint influence in its reports.
Training-free RadAgent already beat CT-Chat on macro-F1; reinforcement learning added further gains and improved external generalisation.
Safety / limitations
Faithfulness of 37% leaves substantial room to improve.
Requires a multi-GPU deployment with heavy tools, which may limit use in resource-constrained settings.
Policy is optimised for a fixed toolbox, so retraining is needed when tools change.
Potential impact: By turning AI-led CT interpretation into a stepwise, inspectable process, RadAgent improves interpretability and offers a more credible path to imaging AI that clinicians can interrogate and regulators can validate.
TOP NEWS
🌰DeepHealth’s Prostate AI gets FDA clearance and CE Mark
Source: DeepHealth · 19 May 2026
DeepHealth builds AI-powered imaging and health informatics tools for radiology, spanning enterprise imaging, population health and disease-specific suites. Its Prostate AI, part of the Prostate Suite, has received FDA 510(k) clearance and CE Marking.
The tool supports prostate MRI interpretation by automating:
Lesion detection
Risk classification
Gland segmentation
PSA density calculation
PI-RADS-compliant reporting
Image source
Why it matters: Prostate MRI interpretation can vary by reader experience and local workflow. In a real-world deployment, DeepHealth says its AI model:
Detected 27% more prostate lesions
Reduced segmentation variability by 65%
Cut workflow time by 37% for biopsy-recommended cases
TOP AI MODEL
Hygieia is an AI model that diagnoses rare diseases and ranks the genes most likely behind them
Source: arXiv (Yale, Duke-NUS, Stanford, Broad and collaborators)
Hygieia tackles two hard problems at once: identifying rare disease and prioritising the genes most likely to explain a patient's phenotype. A KNN-based router decides whether to use a lightweight LLM pipeline for common cases or a fuller agentic loop for rare ones, including knowledge retrieval, patient retrieval, summarisation, and a Claude-Sonnet-4.5 verifier.
The system also estimates confidence through majority voting across repeated runs, which gives clinicians a clearer signal about when the model is more or less certain.
What stands out:
Outperforms o4-mini, GPT-4o, GPT-5 with search, DeepSeek-v3.1, Qwen3-8B, and biomedical agents like Biomni across RareBench splits, MyGene2, RareArena, and in-house Yale data.
Beats certified genetic physicians by 12.49% on rare-disease diagnosis and 60% on risk-gene prioritisation in a head-to-head comparison with the same resources.
Completes the task set in under two hours, versus two to ten hours for human experts.
Produces structured, hierarchical reasoning traces with confidence calibration.
Code released under MIT licence; IRB-approved at Yale.
Caveat: Heavy reliance on closed-source LLM backbones, limited gene-level ground truth, and simulation-derived data temper claims about real-world clinical readiness.
BEDSIDE BETS
Bedside Bets
Startup rounds, deals, and moves.
Blank Bio raises $7.2M to build RNA foundation models for oncology: Built as an applied AI research lab, Blank Bio uses foundation models to learn from tumour RNA data and support biomarker discovery, trial design and clinical diagnostics.
Optura raises a $17.5M Series A, bringing total funding to over $25M: Its enterprise healthcare AI platform helps health plans and providers prioritise AI use cases, turn workflows into AI agents and track return on AI.
Nexalin closes a $1.3M stock-based acquisition of PONM: Best known for non-invasive neurostimulation, Nexalin is adding PONM's AI-integrated remote monitoring, clinical data capture and virtual-care infrastructure to its neurocare platform.
NEWSLETTER BY:
Dr Ezekiel Dinama
MD and PhD Researcher at Cambridge University applying physics-informed ML/AI to neurophysiological research.