Clinical genetics interpretation is not a sequencing problem — it is a meaning problem. A modern whole-genome run can surface four to five million variants per patient, and the overwhelming majority sit in a category labelled variants of uncertain significance. The clinical question is never “what did the sequencer read”; it is “which of these readings should change what we do next Tuesday morning”. That gap, between raw sequence and clinical action, is where AI now does most of its useful work — and increasingly, where computer vision quietly sits alongside the sequence models. We work with research and clinical teams that have already accepted the variant-calling pipeline as solved. What still hurts is the layer above: linking variants to literature, reconciling them with histology and imaging, and producing a report a clinician can sign. The patterns below are what we see when those teams try to industrialise that layer. Why interpretation, not sequencing, is the bottleneck Sequencing cost dropped roughly four orders of magnitude over the last fifteen years (market-direction; not an operational benchmark, but the directional shape is well documented across NHGRI cost curves). Interpretation cost did not. A specialist reviewing a complex panel still spends hours per case reconciling ACMG criteria, family history, ClinVar entries, and primary literature. The labour shape is essentially unchanged from a decade ago. That asymmetry is what AI is now absorbing. The work has three structurally different sub-tasks, and they need different model families: Sub-task Dominant model class What it actually replaces Filter variants by likely pathogenicity Gradient-boosted or deep classifiers over annotated features First-pass tier-1/tier-2 triage Link variants to existing knowledge LLMs + retrieval over ClinVar, OMIM, primary literature Manual evidence assembly Connect molecular findings to images and tissue Computer vision over histology, karyotype, radiology Cross-modality reconciliation Most teams start with the first row, discover the second row is where the time really goes, and only later realise the third row is what makes the report defensible. How does AI actually classify a variant? Variant pathogenicity prediction is now dominated by deep models trained on protein structure, evolutionary conservation, and population frequency. AlphaMissense, ESM-based variant scorers, and SpliceAI are the named tools that recur in the pipelines we see. They do not “decide” — they produce a calibrated score that feeds into the ACMG/AMP framework as one piece of evidence. The honest framing is this: the classifier narrows a five-million-variant haystack to a few hundred candidates, and from there a human geneticist applies the rest of the ACMG criteria. In our experience across biotech R&D engagements with outcome ownership, the realistic compression is roughly two to three orders of magnitude before a human enters the loop — and that compression only holds when the model has been calibrated on a population that resembles the patient. This is an observed-pattern across engagements, not a benchmarked rate against a fixed test set. LLMs and NLP carry the second sub-task. They extract phenotype mentions from clinical notes, link them to HPO terms, and pull supporting evidence from PubMed and case reports. The value is not generation — it is structured extraction at scale. A retrieval-augmented LLM that returns three citations per claim is far more useful in this setting than one that summarises freely. Where computer vision enters a genetics pipeline Genetics is usually framed as a sequence problem, but the clinical record is not purely textual. Karyotypes, FISH images, histology slides, IHC stains, and radiology studies all sit in the same chart. A complete interpretation often requires reconciling a molecular finding with what the tissue or scan actually shows. Computer vision sits in three concrete places in this pipeline: Cytogenetics. Convolutional models segment chromosomes from metaphase spreads and flag structural abnormalities. The throughput gain over manual karyotyping is large; the model still hands ambiguous cases back to the cytogeneticist. Digital pathology. Whole-slide image classifiers identify tumour regions, grade them, and quantify biomarkers. When a somatic variant is called from a tumour sample, the slide-level CV output gives the geneticist a tissue-context check. Radiology-genomics linkage. In oncology especially, imaging features extracted by CV models correlate with molecular subtypes. This is the layer where deep learning in medical computer vision intersects most directly with genomics. The connection back to AI-enabled medical devices is important here. Any CV component that materially influences a clinical decision in the United States is likely Software as a Medical Device (SaMD) under FDA rules. That is not a footnote — it changes how the model is versioned, validated, and updated. Lock-and-key versioning, predetermined change control plans, and post-market surveillance are the operational realities. A CV layer in a genetics pipeline that touches diagnostic output inherits those constraints whether the team designed for them or not. What does production look like when CV and genomics share a pipeline? The integration patterns we see fall into a recognisable shape: Sequencing and imaging arrive on different timelines. A blood draw and a tumour biopsy do not produce results the same week. The pipeline has to tolerate asynchronous arrival and resolve identity correctly. PACS, LIS, and EHR are three separate systems. Imaging lives in PACS, sequence data in the LIS or a bioinformatics platform, and structured clinical data in the EHR. A unified view requires HL7/FHIR connectors, not just a shared dashboard. GPU inference is centralised. Histology slides at gigapixel scale and large variant-scoring models both benefit from shared GPU capacity. Treating GPUs as a per-application resource quickly becomes uneconomic. Audit trails are first-class. Every model output that contributes to a clinical report needs a stored model version, input hash, and timestamp. This is non-negotiable under FDA SaMD expectations and most European regulatory equivalents. Teams that retrofit these properties after a research prototype works tend to lose six to twelve months at the validation stage. Teams that design for them from the start — what we describe more fully in our work on computer vision in biomedical applications — reach a clearable artefact much faster. Generalisability, drift, and population shift The hardest honest claim about AI in clinical genetics is that most published models have been trained on data heavily skewed toward European-ancestry populations. Allele frequencies differ across populations; a variant rare in one group may be common and benign in another. A pathogenicity classifier calibrated on the wrong distribution will over-call variants of uncertain significance in under-represented groups. Drift is the second axis. ClinVar reclassifies variants regularly. A model that scored a variant as likely pathogenic two years ago may need to update when the underlying evidence shifts. The pipeline needs a re-scoring mechanism, and clinical reports issued in the interim need to be re-reachable when a reclassification changes a recommendation. The third axis is technical: sequencing platforms, capture kits, and bioinformatics versions change. A model trained on one upstream stack does not automatically transfer to another. Validation evidence is platform-bound, not just population-bound. We treat these three axes — population shift, evidence drift, technical drift — as the standing risk register for any genetics-CV pipeline we help build. Real-time decision support, honestly framed “Real-time” gets overused. In a newborn intensive care setting, rapid whole-genome sequencing can return a clinically actionable result within hours, and AI-assisted interpretation is a real contributor to that timeline. In a routine outpatient setting, “real-time” means the interpretation arrives during the consult rather than two weeks later — which still matters, but is a different engineering problem. In both cases, the AI layer does not give the final answer. It surfaces the top candidates with their evidence, and the clinician decides. The systems that work are the ones whose interfaces make it easy to disagree with the model: every score links back to the underlying evidence, every flag can be dismissed with a reason, and every dismissal feeds the audit trail. FAQ How many AI-enabled medical devices has the FDA cleared, and which CV patterns recur across them? The FDA’s public list of AI/ML-enabled medical devices has grown into the high hundreds, with radiology dominating the count and ophthalmology, cardiology, and pathology following. The recurring CV patterns are detection (CADe), characterisation (CADx), and segmentation — most cleared devices fall into one of these three categories. What are the production patterns behind FDA-cleared CV diagnostics (CADe, CADx, radiomics)? Cleared devices share a small set of patterns: locked model versions, predetermined change control plans for updates, validation evidence on a defined intended-use population, and post-market surveillance. The model is treated as a regulated artefact, not a continuously updating service. How does deep learning in medical CV translate into regulatory artefacts? Each clinical claim — sensitivity, specificity, time-to-result — must be backed by validation evidence on a representative dataset. The regulatory artefact is the full chain: training data description, model architecture, validation protocol, results on the intended-use population, and a plan for monitoring real-world performance. Where do AI medical-device pipelines need to handle generalisability, drift, and population shift? At three points: training-data composition (population coverage), evidence updates (ClinVar reclassifications, new literature), and technical drift (sequencing platform, capture kit, bioinformatics versions). Each needs a defined monitoring and revalidation mechanism. What integration patterns connect CV inference to PACS, EHR, and clinical workflow? HL7/FHIR for structured clinical data, DICOM for imaging, and direct LIS integration for sequencing. The CV inference service typically sits behind an API that PACS or the EHR calls, with results written back as structured observations rather than free-text reports. Which AI-enabled medical-device companies and products define the current state of practice in 2026? The cleared-device landscape is dominated by radiology vendors (large incumbents and a long tail of specialised startups), with growing presence in digital pathology and ophthalmology screening. The state of practice is defined less by any single product than by the regulatory pattern those products share. How TechnoLynx Can Help We design AI systems for clinical genetics teams that have already accepted the regulatory reality — that any CV or model component touching diagnostic output is a regulated artefact, not a research prototype. Our engagements typically cover the interpretation layer (variant scoring, evidence retrieval, LLM-based summarisation), the CV layer where it intersects with genomics (cytogenetics, digital pathology, radiology-genomics linkage), and the integration layer that connects both to PACS, LIS, and EHR systems under HL7/FHIR. What we bring is the discipline of building for FDA SaMD validation from day one rather than retrofitting it at submission. That includes lock-and-key model versioning, audit trails per inference, predetermined change control plans, and monitoring for population shift and evidence drift. Contact us if you are scoping a genetics-CV pipeline and want the regulatory shape designed in from the start. Image credits: DC Studio and Freepik.
