• Tomorrow's Dose
  • Posts
  • Edition 16 - AI at Work: Differential Performance, Validated Decisions, and Nuclear Medicine's Precision Promise

Edition 16 - AI at Work: Differential Performance, Validated Decisions, and Nuclear Medicine's Precision Promise

Discover how LLMs boost novice MRI readers by 19 percentage points but barely move expert neuroradiologists, explore the first FDA-cleared AI tool for breast cancer treatment selection validated against RCT data at ASCO, and learn how AI-guided patient selection is being positioned to define the next era of radioconjugate therapy in nuclear medicine.

  1. AI LLMs boost novice MRI readers by 19 points — but expert neuroradiologists? Barely.

  2. Artera AI wins FDA clearance for breast cancer treatment decisions — then validates it live at ASCO 2026.

  3. Nucs AI and AstraZeneca partner on AI response prediction for next-generation PSMA radioconjugates in prostate cancer.

  • Featured follow of the week

  • Top posts of the week across social

  • Meet the editor

  • Want a featured article?

Specialty: Radiology // Sub-Specialty: AI // Body Site: Brain

1. AI diagnostic aid helps novice MRI readers, but experts not so much

A multicentre study from Technical University of Munich tested three large language models — GPT-4.1, Gemini 2.5 Pro, and DeepSeek-R1 — across 12 readers at three experience levels, each interpreting 40 confirmed brain MRI diagnoses. Readers first submitted differential diagnoses unassisted, then revised them after LLM-generated suggestions based on their own imaging descriptions. Neurology and neurosurgery residents with no formal radiology training gained an average of 19.4 percentage points in top-three diagnostic accuracy (43.2% to 62.6%; p<0.001), radiology residents gained 14.7 points, but board-certified neuroradiologists improved by only 4.4 points — a difference that did not reach statistical significance. The correlation traced directly to description quality: expert radiologists produced more complete and accurate imaging findings, leaving less room for LLM correction to add value.
Read Full Article

Paul’s Thoughts:

This is one of the most practically important AI papers of the year, and it should change how departments think about deploying LLM-based decision support. The 19-point accuracy gain for non-radiology trainees is significant — but the near-zero gain for experienced neuroradiologists shows that AI isn't a general performance enhancer; it's a gap-filler. At GMI, we've seen this pattern with other AI tools: the populations who benefit most are rarely the ones who drive purchasing decisions. The risk I'd flag is automation bias — if novice readers accept LLM suggestions without critical evaluation, a confident wrong answer from GPT-4.1 could be more dangerous than no suggestion at all. The field needs not just studies showing average accuracy gains, but data on how often LLMs confidently suggest incorrect diagnoses to low-experience readers, and what the clinical consequence is. That's the trial the accompanying editorial rightly calls for.

Timescale: Acute | 1 Year

Specialty: Oncology // Sub-Specialty: AI // Body Site: Breast

2. Artera receives FDA clearance for ArteraAI Breast and presents ASCO 2026 oral validation against SWOG S8814

The US Food and Drug Administration cleared ArteraAI Breast on 6 May 2026, making it the first FDA-cleared multimodal AI (MMAI) tool for HR+/HER2– invasive breast cancer, combining pathology image analysis with clinical and genomic data to predict chemotherapy benefit and recurrence prognosis. The clearance followed prior FDA authorisation of Artera's prostate cancer platform and reflects the company's strategy of building MMAI tools across tumour types validated against randomised controlled trial datasets. At the ASCO 2026 Annual Meeting on 1 June 2026, Artera presented an oral abstract — External Validation of a MMAI Model for Prognosis and Chemotherapy Benefit Prediction in Postmenopausal Node-Positive HR+ Breast Cancer: Analysis of SWOG S8814 (Abstract #107) — extending validation of the MMAI model into the clinically challenging 1–3 node-positive population, where uncertainty about chemotherapy benefit is among the most contested decisions in oncology. The study provides external validation against one of the most important breast cancer RCTs available.

Paul’s Thoughts:

The combination of FDA clearance and live ASCO oral validation at the same time is unusual but important. The SWOG S8814 dataset is a well-respected, hard-won clinical trial resource — using it as an external validation set, rather than internal training data, is the kind of methodological rigour needed. The question I'd push back on is the clinical pathway after clearance: in the US, a cleared tool can be deployed; in the UK and Cyprus, the evidentiary bar is different, and NICE's track record of declining to recommend AI diagnostics without prospective UK data remains a structural barrier. The 1–3 node-positive population is where oncologists carry the most clinical uncertainty, so if this tool performs well in prospective deployment it could genuinely reduce both over- and under-treatment. But "FDA cleared" and "ready for your institution" are not the same sentence — and there is still a gap between them.

Timescale:  Acute | 1 Year

Specialty: Nuclear Medicine // Sub-Specialty: Therapeutics // Body Site: Prostate

3. Nucs AI and AstraZeneca collaborate on AI-driven response prediction for PSMA-targeted radioconjugates in prostate cancer

Nucs AI — a company developing AI-assisted theranostics with a platform built on Lutetium-177 therapy response prediction — announced a collaboration with AstraZeneca on 29 May 2026 to develop a predictive image-based AI algorithm for PSMA-targeted radioconjugate research in metastatic prostate cancer. The aim is to shift patient selection from population-level eligibility criteria toward individual-specific response prediction, tuning models to the specific pharmacological characteristics of next-generation radioconjugates under AstraZeneca's pipeline rather than conventional Lutetium-based therapy. The collaboration positions AI biomarker development to scale in parallel with therapeutic development — a model increasingly favoured by large pharmaceutical companies integrating precision selection into radiopharmaceutical trials.

Paul’s Thoughts:

This collaboration represents a structural shift in how nuclear medicine is approaching patient selection. The conventional Lu-177 PSMA pathway already has strong theranostic logic — image it with PSMA PET, treat it with PSMA-targeted Lu-177 — but the binary nature of this selection leaves a significant population in a grey zone where current eligibility criteria are imprecise. Shifting to AI-driven individual response prediction is the right direction, and AstraZeneca's involvement is a credibility signal. The practical constraints are data access and regulatory pathways for AI-guided trial enrolment in Europe. At GMI, our SubtlePET experience has shown us that AI in nuclear medicine can work extremely well at the image-processing layer; the harder problem is what happens when AI gates access to therapy itself. That's a clinical governance question the field is not yet ready to answer, but this collaboration will force it.

Timescale: Early | 2 Years

Prof. Dr. Lena Maier-Hein Headshot

Prof. Dr. Lena Maier-Hein

Prof. Dr. Lena Maier-Hein Head of Intelligent Medical Systems, German Cancer Research Center (DKFZ) | Managing Director, National Center for Tumor Diseases (NCT) Heidelberg | Professor, Heidelberg University

Follow Prof. Maier-Hein for one of the most rigorous European perspectives on the validation and clinical translation of AI in medical imaging — her work focuses on how to build algorithms that can be trusted in practice, not just benchmarked in research, and she leads the DKFZ's Intelligent Medical Systems division where the gap between AI promise and clinical deployment is addressed head-on. This week's edition connects directly to her long-running agenda: Story 1's finding that LLM benefit depends on reader quality, and Story 2's emphasis on RCT-validated AI deployment, are precisely the translation and validation challenges that drive her research programme.

Linkedin logo button follow

A round-up of some of the best posts we found online this week.

Was this email forwarded to you?
Our weekly email brings you the latest health trends and insights, combining top news and opinions into a straightforward, digestible format.

Want an article featured?

Have an insightful link or story about the future of medical health? Reach out below, and we may include it in a future release.

Reply

or to participate.