Clinical AI · Chest Radiography · Computational Biology
A curated overview of how AI is used to analyze chest radiographs and MRI, work with DICOM studies, support first-line screening in ordinary clinics and design new therapeutic proteins and antibodies.
AI + chest radiography
Chest X-ray is cheaper, faster and far more widely available than CT. That makes it attractive for mass checkups, occupational screening, routine dispensary programs and a first visit in ordinary polyclinics. AI can help by flagging suspicious findings on plain radiographs so that patients who need closer review or CT can be prioritized earlier.
On a standard chest radiograph, subtle nodules or focal opacities can be easy to miss, especially in busy first-line care. AI systems can scan the image, highlight suspicious regions and support a faster referral pathway for radiologist review, repeat imaging or CT confirmation.
Illustrative demonstration figure for educational use.
For MRI, segmentation models can outline tumor boundaries and enable volume tracking over time. This is a different use case from population screening, but it shows how AI ranges from simple frontline triage on X-rays to detailed quantitative support on complex scans.
Illustrative demonstration figure for educational use.
DICOM in practice
DICOM carries pixel data together with structured metadata that identifies the study, series, modality, acquisition parameters and spatial context. That matters not only for CT and MRI, but also for digital radiography workflows when AI results need to return to PACS cleanly and consistently.
Read the current DICOM PS3.18 standard ↗
A chest X-ray, CT or MRI study is stored in PACS/VNA with DICOM metadata.
Series selection, resampling, normalization and quality checks.
The model produces detections, masks, measurements or probabilities.
Results return to the workflow as overlays, structured data or triage signals.
Generative AI + Biologics
For therapeutic antibodies, computational models are being used for structure prediction, epitope targeting, affinity optimization, developability screening and de novo generation. Experimental validation is still essential: a model proposal is not a drug.
A 2025 Nature study reported computational generation of antibody formats aimed at user-specified epitopes, followed by experimental screening and structural characterization. It is a concrete example of generative protein design moving beyond sequence ranking.
Nature paper ↗Absci has publicly described ABS-501 as an AI-designed anti-HER2 antibody program. The company reports preclinical leads with novel epitope interactions and activity in a trastuzumab-resistant xenograft model. These are company-reported preclinical data, not evidence of clinical efficacy.
Company update ↗A 2025 review in npj Precision Oncology describes AI applications spanning tumor-target selection, antibody engineering, linker/payload optimization, ADMET modelling and patient stratification.
Open-access review ↗Research & signals
Selected links worth following if you want to see the field evolve rather than just read summaries about it.
Regulatory database with device, company, panel and decision data.
npj Precision Oncology · 2025AI in lung cancer clinical translationOpen-access review covering prevention, screening, diagnosis, prognosis and treatment — useful context for chest X-ray and lung cancer workflows.
X · Radiology: AIAgentic AI in radiologyLink to a 2026 discussion of LLMs and future clinical integration.
X · RadiologyClinical imaging AI discussionAn example radiology-community post showing how imaging AI is discussed outside formal journal articles.
X · ESRAI/ML demos at ECR 2026A snapshot of emerging imaging software presented to the radiology community.
Nature · 2025AI agents design nanobodiesThe Virtual Lab study combined AI agents and computational biology tools with experimental validation.
npj Digital Medicine · 2026Generative antibody engineeringRecent perspective on the shift from empirical screening toward specification-driven design.
Responsible clinical AI
Performance should be tested beyond the development dataset, ideally across sites, scanners and patient populations relevant to intended use.
A model can have high discrimination and still produce poorly calibrated probabilities or systematic errors in important subgroups.
The way an alert is shown, escalated and documented matters, especially when AI is used at scale in routine outpatient screening rather than in a single specialist center.
About this project
AI in Medicine collects examples, papers and technical concepts at the intersection of medicine and artificial intelligence. It does not provide medical advice, diagnosis or treatment.