Clinical trials run on imaging endpoints, paper trails, and site logistics — three streams where computer vision now does work that used to absorb coordinator time. The interesting question is not whether CV belongs in trials (it already does, across imaging core-labs and eCOA workflows), but where it earns its place against the regulatory and operational constraints of a study under protocol. This is a supporting view to our hub on AI-enabled medical devices and the computer vision layer behind FDA-cleared tools. The hub explains the SaMD frame; this article narrows in on how CV shows up inside an active trial — what it actually accelerates, what it cannot touch without revalidation, and what the integration surface looks like. What computer vision does inside a trial A clinical trial has three streams where image and video data accumulate faster than humans can process them: Imaging endpoints. MRI, CT, ultrasound, OCT, and digital pathology scans feed primary or secondary endpoints. Convolutional neural networks (CNNs) measure lesions, segment regions of interest, and track change over time. In oncology trials this work flows through a central imaging core-lab; the CV layer is a measurement instrument, not a diagnostic device. Trial documentation. Consent forms, source documents, shipment slips, and lab notes still arrive as scans or phone photos. Optical character recognition (OCR) lifts the text out, and structured field extraction maps it to the trial database. Site operations. Kit inventory, visit-room setup, sample-handling steps, and device check-ins generate camera streams that object-detection models can read in near real time. These are different problems with different evidence bars. The imaging-endpoint stream sits closest to the regulator; the documentation and site-ops streams sit closer to operational efficiency and GCP audit-readiness. Treating them as one bucket — “AI in trials” — is the easiest way to misjudge what each one actually needs. How the imaging-endpoint layer differs from the rest This is the distinction practitioners care about, and it is worth being explicit. Stream Regulatory weight Validation evidence Change-control friction Imaging endpoints (core-lab measurement) High — feeds primary/secondary endpoints; subject to imaging charter and central read Locked model version, reader-agreement studies, bias analyses across scanners and sites High — model updates trigger re-read or sensitivity analysis Trial documentation (OCR + extraction) Low–medium — supports source-data verification, not the endpoint itself Field-level accuracy on representative document sets; human-in-the-loop reconciliation Low — extractor can be updated within GCP change-control Site operations (object detection, kit tracking) Low — operational signal, not trial data Operational accuracy on site-specific footage; thresholds tuned per site Low — local retraining is acceptable The pattern we see across engagements is that teams underestimate the first row and overestimate the third. The endpoint layer needs lock-and-key model versioning of the kind described in the hub on FDA-cleared CV tools; the site-ops layer can iterate on weekly cycles. Conflating the two — applying medical-device discipline to a kit-counting model, or applying ops-grade iteration to an endpoint reader — costs months either way. Imaging endpoints: where CV is doing measurable work Quantitative imaging biomarkers (QIBs) — tumour-volume change, ejection fraction, retinal-layer thickness, white-matter lesion load — are increasingly produced by CNN segmentation rather than manual delineation. In our experience across imaging core-lab engagements, the operational gain is not raw speed (a single reader can be fast) but consistency: model output is deterministic given the same input, so inter-reader variance drops to the variance the model itself was validated against. That is what matters for endpoint power calculations. The constraint is generalisability. A segmentation model validated on Siemens 3T scans may underperform on GE 1.5T data from a site that joined late. This is the population-shift problem discussed in the hub; inside a trial, it shows up as scanner-stratified bias analysis in the statistical analysis plan. Tooling here is mature: nnU-Net and MONAI for segmentation, DICOM-native I/O via pydicom and dcmqi, ONNX for deployment into the core-lab’s reader workstation, and CUDA/TensorRT for the inference path when read volumes justify it. None of this is exotic; what is exotic is the documentation burden around it. OCR and document extraction The trial documentation stream is where CV pays back fastest and least visibly. Consent forms, lab requisitions, IRT shipment confirmations, and source documents from non-electronic sites still arrive as PDFs and phone photos. A modern OCR pipeline — Tesseract for printed text, a transformer-based layout model (LayoutLMv3 or Donut) for structured forms, and a downstream extractor that maps fields to CDISC-aligned variables — can collapse days of data-entry work to minutes. Two practical notes from engagements we have seen: Field-level accuracy, not document-level accuracy, is the metric. A 98% page-level “looks right” score hides the 2% of missing batch numbers that block data lock. Human-in-the-loop reconciliation is mandatory at this stage of maturity. OCR for trial documents augments coordinators; it does not replace source-data verification. Site operations and kit logistics This is the lowest-stakes, highest-volume layer. Cameras over packing benches and visit-kit shelves, paired with an object-detection model (a YOLO-family detector is the common choice; DETR variants when the scene is cluttered), produce kit counts that reconcile against IRT shipment data. The model does not need to be brilliant; it needs to be reliable under the lighting and angle of the actual room. What usually fails here is not the model but the camera placement. A site that moves a shelf two metres breaks the field-of-view assumption, and the model starts under-counting. The fix is procedural — a site-acceptance checklist for any CV-monitored space — not algorithmic. Why “computer vision in trials” is not one project The throughline across all three streams is that none of them is generic AI work. Each one has its own evidence bar, its own integration surface (PACS and DICOM for imaging; EDC and CDISC for documents; IRT and inventory systems for ops), and its own change-control regime. We see programmes stall when they try to procure “an AI vendor for the trial” as a single line item; we see them move when the three streams are scoped, governed, and budgeted separately. For the regulatory frame around the imaging-endpoint stream — locked model versioning, validation evidence, the SaMD distinction — the hub on FDA-cleared CV tools is the reference. For the deep-learning architectures behind the segmentation work, our companion article on deep learning in medical computer vision covers the model side in more depth. FAQ How many AI-enabled medical devices has the FDA cleared, and which CV patterns recur across them? The FDA maintains a public list of AI/ML-enabled devices; radiology dominates, followed by cardiology and ophthalmology. The recurring CV patterns are segmentation for quantitative biomarkers, detection for triage (CADe), and classification for screening (CADx). The hub article walks through the categories in detail. What are the production patterns behind FDA-cleared CV diagnostics (CADe, CADx, radiomics)? A locked model version, a documented training and validation dataset, reader-agreement studies against expert ground truth, and a post-market surveillance plan. Production deployment is typically through a PACS-integrated reader application using DICOM I/O; inference may be on-prem (regulated environments) or in a validated cloud region. How does deep learning in medical CV (classification, segmentation, detection) translate into regulatory artefacts? Each model task maps to specific validation evidence. Classification needs ROC/AUC and operating-point analysis; segmentation needs Dice/Hausdorff against expert masks and inter-rater comparison; detection needs sensitivity/specificity at clinically meaningful thresholds. All three need bias analyses across demographics and acquisition equipment. Where do AI medical-device pipelines need to handle generalisability, drift, and population shift? Generalisability is handled at validation: stratified evidence across scanner vendors, field strengths, demographics, and clinical settings. Drift is handled post-market: monitoring of output distributions against the validation reference, and a documented retraining trigger. Population shift is the highest-risk variant — a model deployed at sites outside the validation population needs a sensitivity analysis before its outputs feed endpoints. What integration patterns connect CV inference to PACS, EHR, and clinical workflow? DICOM SR (Structured Reporting) for measurements back into PACS; HL7 FHIR for EHR-side observations; vendor-neutral archives (VNAs) for image routing. The integration is more often the project bottleneck than the model itself. Which AI-enabled medical-device companies and products define the current state of practice in 2026? The current state of practice is defined by the FDA’s published AI/ML device list. Rather than name vendors here, we point readers to that public list and to the hub article for the categories where clearance density is highest. How TechnoLynx works on trial CV We build computer vision pipelines for clinical-trial sponsors and imaging core-labs. Our work covers segmentation for quantitative imaging endpoints (nnU-Net, MONAI, ONNX deployment), OCR and structured extraction for trial documentation, and operational object-detection layers for site logistics. We design for the regulatory bar of each stream — endpoint-grade evidence for the imaging layer, GCP change-control for documents and operations — and we integrate against PACS, EDC, and IRT rather than asking sites to adopt a new tool. If you are scoping a CV layer into an upcoming protocol or imaging charter, the conversation starts with which of the three streams you are actually trying to instrument, and what evidence the regulator and your statistical team will accept.
