Computer Vision for Medical Imaging: AI Diagnostic Support for Radiology and Pathology Teams

General information only — not medical advice. AI systems for medical imaging are regulated medical devices under EU law. Always involve qualified clinicians and a notified body before deploying any diagnostic AI in a clinical setting.

AI medical imaging diagnostic support has moved from academic research into live clinical pilots faster than almost any other AI application in healthcare. Radiology departments across the Netherlands are evaluating — or already running — deep-learning tools that scan MRI, CT and X-ray images for anomalies before a radiologist opens the worklist. Pathology labs are piloting whole-slide image analysis that flags suspicious tissue regions in digitised biopsy slides. Dermatology services are trialling image-recognition tools that pre-screen lesion photographs. None of these tools diagnose on their own — they are decision-support systems that assist trained clinicians. But when they are designed, validated and integrated properly, they can help hard-pressed departments handle rising scan volumes without compromising clinical quality.

This guide explains how computer vision for medical imaging works in practice, what the regulatory environment demands, where the technology genuinely helps, and how Crux Digits supports Dutch healthcare organisations that want to explore or implement these capabilities responsibly.

What AI tools support medical image analysis for diagnostic teams in the Netherlands?

This is the question Dutch radiology heads, clinical informatics leads and hospital innovation managers are asking most frequently in 2026. The answer covers several distinct tool categories, each operating at a different point in the clinical workflow.

Radiology triage and worklist prioritisation

The most widely deployed category today is AI-assisted triage: a model processes an incoming scan immediately after acquisition and assigns a flag — often a severity score or a binary urgent/non-urgent label — before the scan reaches the radiologist worklist. If the model detects features consistent with intracranial haemorrhage on a CT, for example, it pushes that scan to the top of the queue so the most time-sensitive cases are reviewed first.

These triage tools do not produce a diagnosis. They re-order the worklist. That distinction matters legally and clinically. The radiologist still reads every scan; the tool simply changes the sequence in which they are read. Under EU regulatory frameworks this is still considered a medical device function — but the risk profile is different from a tool that generates a primary diagnosis, and validated triage assistants from CE-marked vendors are now in operational use in Dutch hospitals.

Anomaly detection and AI-assisted reading in radiology

Computer vision radiology AI Netherlands deployments most frequently target high-volume, pattern-recognition-heavy workflows: chest X-ray screening for tuberculosis, pneumonia and nodules; mammography CAD (computer-aided detection) for microcalcifications; CT colonoscopy polyp flagging; and lung nodule detection in low-dose CT screening programmes. In each case the model analyses the image and produces a set of region-of-interest annotations — bounding boxes, heatmaps or segmentation masks — overlaid on the image for the radiologist to accept, reject or modify.

The radiologist remains the reporting physician. The AI annotation is a prompt, not a finding. In well-designed systems the radiologist can see how confident the model is and can dismiss annotations they consider false positives with a single click. The interaction is designed to assist, not to override.

AI-assisted pathology image analysis

AI-assisted pathology image analysis operates on digitised whole-slide images (WSI) — high-resolution scans of tissue sections stained on glass slides. A deep-learning model trained on annotated pathology slides can scan a full WSI in minutes and highlight regions that contain cells exhibiting specified morphological features: nuclear atypia, mitotic figures, tumour-infiltrating lymphocytes, glandular architecture patterns relevant to Gleason grading in prostate cancer, or invasive front characteristics in colorectal specimens.

Dutch pathology laboratories that have adopted digital pathology platforms — a requirement the field is moving towards for both quality and workflow reasons — are well-positioned to add AI analysis layers. The challenge is not principally the AI itself; it is the image management infrastructure, the labelling of ground-truth slide annotations from specialist pathologists, and the validation studies required before clinical use. Crux Digits supports the engineering side of this stack: ingesting WSI data, building and validating the deep-learning pipeline, and integrating outputs into the laboratory information system.

Deep learning medical scan analysis: MRI and CT

Deep learning medical scan analysis on volumetric imaging — 3D MRI and CT datasets — presents a different set of engineering challenges from 2D classification tasks. A single abdominal CT may contain 400-600 slices; a brain MRI protocol may involve multiple sequences each with hundreds of slices. Models must reason across the full volume, not just individual frames, to detect subtle findings that only become apparent in three-dimensional context.

Common applications include: automated measurement of organ volumes and lesion dimensions for longitudinal follow-up; segmentation of brain structures to assist neurological assessment; bone and joint analysis in musculoskeletal imaging; and vertebral fracture detection in spine CT. These tools generate structured outputs — measurements, segmentations, flags — that can feed directly into structured radiology reports, reducing the manual data-entry burden on reporting radiologists.

AI dermatology image recognition

AI dermatology image recognition was one of the earliest clinical imaging applications to attract serious research attention, and it has matured considerably. Convolutional neural networks trained on large dermoscopy datasets can classify lesion images against a taxonomy of common skin conditions — melanoma, basal cell carcinoma, seborrhoeic keratosis and others — with performance benchmarks that have been extensively published in peer-reviewed literature.

In clinical practice, dermatology AI tools are most valuably deployed in triage contexts: a primary care physician photographs a concerning lesion, the tool provides a preliminary risk stratification, and the output is used to prioritise referrals to a specialist rather than to make a treatment decision. The tool supports the clinical pathway; it does not replace the dermatologist.

Automated anomaly detection in medical imaging

Automated anomaly detection medical imaging tools are increasingly used in preventive and screening contexts where the base rate of positive findings is low and human fatigue is a genuine quality risk. In a lung cancer low-dose CT screening programme, for example, the vast majority of scans will be negative; radiologist attention naturally varies over a long session of predominantly negative reads. A model that flags potential nodules — even with a high false-positive rate — ensures that the radiologist attention is directed to the regions that warrant it.

The same principle applies in retinal screening for diabetic retinopathy, where graders review large numbers of fundus photographs, and in dental panoramic radiograph analysis for caries and bone-level assessment. These are use cases where AI assistance has clear workflow and quality rationale, and where CE-marked tools are beginning to reach the Netherlands market.

The EU regulatory framework: EU AI Act, EU MDR and IVDR

Any team in a Dutch hospital considering AI for medical imaging needs to understand the regulatory landscape clearly. There are two intersecting frameworks.

EU MDR and IVDR: medical device regulation

Under Regulation (EU) 2017/745 — the EU Medical Device Regulation (MDR) — software that is intended to be used for a medical purpose, including supporting diagnosis, is classified as a medical device and must bear CE marking before it can be placed on the EU market. Software that provides diagnosis or therapy decisions is typically classified as Class IIa or higher, requiring assessment by a notified body.

In vitro diagnostic software — including AI tools applied to pathology images for diagnostic purposes — falls under Regulation (EU) 2017/746 (IVDR), which has similarly demanding conformity requirements. Hospitals procuring AI diagnostic tools must verify CE marking status and scrutinise the intended use statement carefully: a tool whose intended use is described as research-only or decision-support-only-not-for-diagnosis has a different regulatory status from one that carries CE marking as a diagnostic device, and the clinical governance implications differ accordingly.

EU AI Act: high-risk classification

The EU AI Act, which entered application in 2024, classifies AI systems used in medical devices as high-risk under Annex III. This means that — in addition to MDR/IVDR requirements — operators deploying such systems must maintain technical documentation, conduct conformity assessments, implement human oversight measures, register systems in the EU database, and maintain logs of system operation. For hospital innovation and procurement teams, this creates a documentation burden that must be factored into project timelines and governance structures.

Crux Digits works with healthcare clients to navigate both frameworks. We do not certify medical devices — that is the role of notified bodies and the vendor — but we can help healthcare organisations understand which AI applications require MDR CE marking, structure their procurement due diligence, and design the human-oversight workflows and audit trails that both frameworks require.

Why clinical validation is the central challenge

The most common reason AI medical imaging tools fail to move from pilot to production in Dutch hospitals is not technical performance — it is the absence of adequate clinical validation in the local context.

A model may have been trained on a large international dataset and may have published performance benchmarks from a multi-centre study. But if its training data did not include images from your CT scanner model, your staining protocol for pathology slides, or your patient population demographics, its real-world performance in your department may differ materially from the published figures. Dutch hospitals that have conducted rigorous local validation studies — comparing model output against a panel of expert clinicians on prospectively collected local data — have consistently found performance variation compared with published benchmarks, in both directions.

This does not mean published benchmarks are worthless. It means that local validation is an essential step, not an optional one, before clinical deployment. It also means that the data-engineering infrastructure to collect, annotate and manage local validation datasets — and to continue monitoring model performance after deployment — is a first-class project requirement, not an afterthought.

Crux Digits brings data engineering and pipeline expertise to support the infrastructure side of this validation work: ingesting imaging data from PACS, managing DICOM metadata, building annotation workflows with clinical labellers, and constructing evaluation frameworks that give clinical governance committees the evidence they need.

How computer vision pipelines for medical imaging are built

A production computer-vision pipeline for hospital radiology or pathology is considerably more complex than a standalone model. The components that matter in practice are worth describing clearly.

DICOM ingestion and image pre-processing

Medical imaging data is stored in DICOM format, which encodes not just the pixel data but a rich set of metadata: scanner parameters, patient demographics, acquisition protocol, institution identifiers. A robust ingestion layer must extract the relevant pixel arrays, apply protocol-appropriate pre-processing (windowing for CT, normalisation for MRI, colour normalisation for pathology slides), and — critically — de-identify data appropriately before it is passed to any model that was not trained on the original institution data. GDPR and hospital data-governance policies set the rules here, and the engineering must comply from the start.

Model architecture and training

For 2D classification tasks (chest X-ray, dermatology, WSI tile classification) convolutional neural networks and vision transformers are the established architectures. For volumetric tasks (3D MRI/CT segmentation) 3D convolutional architectures or hybrid 2D-slice approaches with sequence modelling are common. Transfer learning from large medical imaging pre-trained models — such as those trained on large radiology report and image corpora — has improved sample efficiency considerably, meaning reasonable performance is achievable with fewer locally annotated cases than was required three years ago.

Crux Digits is vendor-neutral on model architecture. The right choice depends on the imaging modality, the target task, the available annotated data volume and the inference latency requirements. What we bring is the computer vision engineering to translate a clinical problem statement into a model specification, training pipeline and evaluation framework.

Human-in-the-loop annotation workflows

Ground-truth annotation for medical imaging requires clinical expertise. A radiologist must annotate nodules on CT. A pathologist must label glandular regions on a prostate biopsy slide. An annotation tool that works for a general computer-vision dataset is not adequate for this task: it must support DICOM viewing, 3D annotation contexts, structured label taxonomies, inter-annotator agreement tracking and clinical review workflows.

Building and managing this annotation infrastructure is frequently underestimated in medical AI projects. Crux Digits treats it as a first-class engineering deliverable, not a manual data-collection side task. The quality of the annotation directly determines the quality of the model.

Integration with PACS, RIS and LIS

A model that produces good outputs in a research environment but cannot be accessed from the radiologist workstation, or whose outputs do not appear in the radiology information system (RIS) alongside the images, will not be used. Integration with existing PACS (Picture Archiving and Communication System), RIS and laboratory information system (LIS) infrastructure is a core engineering requirement. This typically involves HL7 FHIR or proprietary DICOM DIMSE/DICOMweb interfaces, and the integration design must be agreed with the hospital IT and clinical informatics teams from the start of the project.

Our AI implementation practice covers this integration layer explicitly — from architecture design through to testing, go-live support and post-deployment monitoring.

Monitoring and continuous quality assurance

A deployed medical AI model is not a fixed product. Scanner upgrades, protocol changes, patient population shifts and vendor software updates can all affect model performance. A production deployment needs a monitoring framework: regular performance audits against a reference set of clinician-labelled cases, drift detection for changes in input data distribution, and a governance process for deciding when retraining or recalibration is needed.

This is not optional under the EU AI Act requirements for high-risk AI systems. It is a documented obligation. Building the monitoring infrastructure as part of the initial deployment — rather than retrofitting it later — is substantially more efficient and avoids the compliance gap that leaves some hospital AI deployments in a regulatory grey zone.

Crux Digits and Dutch healthcare: our approach

Crux Digits is a vendor-neutral AI consultancy based in Utrecht, working with Dutch and EU organisations on AI strategy, implementation and data engineering. Within healthcare, our focus is on supporting clinical and informatics teams who want to deploy AI responsibly — with rigorous validation, proper regulatory awareness and human-oversight workflows built in from the start.

We do not build medical devices. We support the engineering infrastructure around them: the data pipelines that feed them, the annotation workflows that produce training and validation data, the integration layers that connect them to clinical systems, and the monitoring frameworks that keep them performing as intended. Where a hospital is evaluating a CE-marked commercial AI product rather than building a bespoke model, we can support the evaluation, the integration and the governance framework.

Our computer vision and AI implementation services are described on our site. For healthcare organisations at an earlier stage — still assessing whether and how to approach medical imaging AI — our advisory engagements begin with a structured discovery session that maps current imaging workflows, identifies the highest-value AI assistance opportunities, and produces a realistic assessment of the data, regulatory and integration requirements involved. See our transparent pricing page for engagement structures, browse our case studies, or get in touch to arrange a no-obligation conversation.

A practical checklist: is your department ready for medical imaging AI?

• Clinical problem definition: Have you identified a specific, narrow workflow problem — triage prioritisation, annotation assistance, measurement automation — rather than a general aim to introduce AI? Specificity drives successful projects.
• Data availability: Do you have access to a sufficient volume of locally labelled imaging data, or a plan to create it? Annotation by qualified clinicians is typically the longest lead-time item in any medical imaging AI project.
• Regulatory status of the tool: If you are procuring a commercial product, does it carry CE marking under MDR or IVDR for the intended clinical use? If you are building bespoke, have you assessed the regulatory pathway?
• EU AI Act compliance: Medical AI is high-risk under the EU AI Act. Does your procurement and governance process include the required documentation, human oversight provisions and registration obligations?
• Integration readiness: Can model outputs be surfaced in the radiologist or pathologist existing workstation workflow? Has your clinical IT team assessed the PACS/RIS/LIS integration requirements?
• Human-oversight design: Who reviews model flags before they influence clinical action? Is this workflow documented and tested?
• Validation plan: Have you defined how you will conduct local clinical validation before go-live, and what performance thresholds are clinically acceptable?
• Post-deployment monitoring: Is there a governance process and technical infrastructure for ongoing performance auditing after deployment?

Medical imaging AI is not a shortcut and it is not a silver bullet. But for departments under real volume and throughput pressure, well-designed AI decision support — implemented with clinical rigour, proper regulatory compliance and sustained human oversight — can make a meaningful difference to how teams work and to the consistency of image review quality. The projects that deliver sustained value are those where the clinical problem is clearly defined, the validation is thorough, the integration is genuinely usable, and the human-oversight workflow is designed in from the start, not bolted on at the end.

Interested in exploring what responsible medical imaging AI could look like for your radiology or pathology department? Contact Crux Digits — we are glad to have a frank conversation about what is feasible, what the data and regulatory requirements look like, and what a realistic implementation path involves.

Frequently asked questions