Computer Vision in Healthcare: Working vs Hype

Healthcare is one of the most promising — and most overhyped — domains for computer vision. Having built CV systems for healthcare clients, we've seen firsthand which applications deliver real clinical value and which remain perpetually "almost ready." Here's an honest assessment.

What's Actually Working in Production

Radiology: Triage and Second-Read Assistance

The most mature healthcare CV application. AI systems that flag urgent findings on chest X-rays, CT scans, and mammograms are deployed across thousands of facilities. These systems work because:

• The images are standardised (DICOM format, consistent acquisition protocols)
• Large labelled datasets exist from decades of radiological reporting
• They augment rather than replace radiologists — flagging priority cases rather than making diagnoses
• Regulatory pathways (FDA 510(k) for triage tools) are well-established

The key insight: the most successful radiology AI doesn't try to replace the radiologist. It helps them see more patients faster by sorting cases by urgency and highlighting regions of concern.

Pathology: Quantitative Analysis

Digital pathology has reached an inflection point. Whole-slide imaging scanners now produce gigapixel images that can be analysed computationally. Working applications include:

• Cell counting and classification: Automating tedious quantitative tasks that are error-prone when done manually under a microscope for hours.
• Tumor margin assessment: Identifying the boundaries of cancerous tissue in surgical specimens to guide treatment decisions.
• Biomarker quantification: Measuring protein expression levels (e.g., HER2 scoring in breast cancer) with greater consistency than manual assessment.

Dermatology: Screening and Monitoring

Skin lesion classification has reached dermatologist-level accuracy in controlled studies. Mobile apps can now photograph a mole and provide a risk assessment. The practical applications:

• Primary care screening to determine which patients need dermatologist referral
• Remote monitoring of chronic skin conditions
• Wound measurement and healing assessment

Ophthalmology: Diabetic Retinopathy Screening

One of the earliest FDA-approved autonomous AI diagnostic systems (IDx-DR) screens for diabetic retinopathy. It works because retinal imaging is highly standardised and the clinical need is clear: millions of diabetic patients need annual screening, and there aren't enough ophthalmologists.

What's Promising But Not Ready

Surgical AI and Real-Time Guidance

Computer vision during surgery — identifying anatomical structures, tracking instruments, providing real-time guidance — is technically impressive but faces major hurdles:

• Extreme reliability requirements (a crash during surgery is not acceptable)
• Enormous variability in surgical anatomy and technique
• Regulatory barriers are much higher for intraoperative tools
• Liability questions remain unresolved

Current applications are limited to post-operative video analysis and training tools rather than real-time guidance.

Mental Health Assessment from Video

Detecting depression, anxiety, or cognitive decline from facial expressions and movement patterns is scientifically interesting but ethically fraught and technically unreliable in diverse populations. The risk of bias across demographics is severe, and the clinical utility compared to established assessment methods is unproven.

Technical Challenges Unique to Healthcare CV

• Data scarcity: Despite healthcare generating massive data volumes, labelled datasets for specific conditions are small. A rare cancer might have only 200 labelled examples available for training.
• Regulatory requirements: FDA approval (or CE marking in Europe) adds 12-24 months and significant cost to any clinical CV product. Your model needs to work across diverse populations and imaging equipment.
• Integration with clinical workflows: The best model in the world is useless if it doesn't integrate with the hospital's PACS (Picture Archiving and Communication System) and EHR (Electronic Health Record). This integration work often exceeds the ML development effort.
• Explainability: Clinicians won't trust a black box. Attention maps, GradCAM visualisations, and other explainability tools are essential for clinical adoption, not nice-to-haves.
• Data privacy: HIPAA compliance, de-identification pipelines, and secure training environments add complexity and cost that don't exist in other CV domains.

Our Approach to Healthcare CV Projects

When we engage on a healthcare computer vision project, we follow a specific protocol:

• Regulatory strategy first. Before writing a line of code, we map out the regulatory pathway. This determines the development process, documentation requirements, and testing rigor.
• Clinical validation design. Working with clinical partners to design validation studies that will satisfy both regulators and clinicians.
• Bias testing. Systematic evaluation across demographic groups, imaging equipment manufacturers, and clinical settings. This isn't optional.
• Integration planning. Understanding the clinical workflow and designing the AI to fit seamlessly into existing processes rather than requiring workflow changes.

Healthcare CV is one of the highest-impact applications of AI, but it demands more rigor, more patience, and more domain collaboration than any other sector. The teams that succeed are the ones who respect the complexity rather than trying to move fast and break things.