Chronic wounds (diabetic foot, pressure and venous ulcers) are measured with rulers and described in free text, so results vary between nurses and healing progress is hard to compare. AI-assisted apps take a photo with a smartphone, calibrate scale, segment the wound, classify tissue and write a structured, comparable record into the care documentation. This is a regulated medical-device function, not a generic measurement tool.
Problem & value
- Today: nurses in home care, nursing homes and wound clinics measure length/width with a paper ruler, estimate tissue colours by eye and write free-text notes. Inter-rater variability is high, photos often sit unstructured in messaging apps or on private phones (a data-protection problem), and stagnation is noticed late.
- AI value: consistent, repeatable area and tissue measurement; time-stamped photo series; structured documentation that satisfies expert standards (e.g. German DNQP Expertenstandard on chronic wounds, international diabetic-foot and pressure-injury guidelines) [1][2]; earlier escalation to a physician or wound specialist; support for telemedicine wound consults.
- Benefit figures: vendors publish clinical validation studies and time-saving claims; treat them as vendor claims and verify against peer-reviewed studies on your own wound mix. No generic figure is given here.
Variants
- Home care (ambulante Pflege): nurse app on a smartphone, offline-capable, sync to care software.
- Nursing homes / long-term care: wound documentation integrated with care planning, see HLTH-025.
- Wound clinics and hospitals: specialist workflow, sign-off by wound experts, link to EHR.
- Telemedicine: photos from general practitioners or patients for remote review by wound specialists.
- Patient self-monitoring (diabetic foot): consumer-grade photo checks; higher regulatory bar and safety risk.
- Veterinary: out of scope, see HLTH-016.
How it works
- capture (smartphone, reference marker or depth sensor)
- quality check (focus, lighting, angle, wound in frame)
- calibration (marker / LiDAR / stereo depth
- real-world scale)
- wound segmentation + tissue classification (granulation, slough, necrosis, epithelium)
- measurement (area, length, width, perimeter; depth/volume if 3D data)
- comparison with previous visits
- trend / stagnation flag
- nurse review and correction (human-in-the-loop)
- structured record to EHR / care software (+ optional LLM-drafted narrative)
- Segmentation and tissue classification: CNN/transformer segmentation models trained on expert-annotated wound images; outputs are pixel masks converted to physical units via calibration.
- Calibration: printed reference stickers, or depth sensors on recent phones/tablets; without it, area is unreliable.
- Human-in-the-loop: the nurse confirms or edits the outline and tissue labels before saving; clinical decisions (treatment, referral, infection suspicion) stay with qualified staff. The AI documents; it does not diagnose infection.
- LLM layer (optional): drafts a text description from structured values; never inserts findings that are not measured.
Data preparation
Data are wound photos, measurements, wound-care plans and patient master data. In practice photos sit on private phones, in messaging apps, on file shares or as attachments in the patient record; measurements are on paper charts or in free-text notes.
Typical gaps: no scale marker or inconsistent distance and lighting, photos without date or wound ID, no patient consent for photography, mixed storage on personal devices, and unclear access rights.
Preparation steps: define a photo protocol (colour/scale reference, distance, lighting, one photo set per visit); move photos from private devices into the record system or the vendor app (PREP-007, PREP-008); add metadata such as wound location, date and etiology (PREP-010); document consent; and, for own models, have wound-care specialists label tissue types and borders with a fixed scheme (PREP-019, PREP-018). Pseudonymise exports for training (PREP-022).
Tools: the product's capture app or a standard camera app with a calibration sticker, CVAT or Label Studio for labelling, the clinic information system for storage. Expect roughly 25-35 percent of the effort for data preparation on path A, more on path C. See GUIDE-003 and PREP-024 for the rule on what may go into which tool.
Implementation paths
Which path for whom
- enterprise: A - certified wound-documentation apps exist for clinics and care providers, and measurement and classification count as a medical-device function.
- developer: C - open segmentation models fine-tuned on a single GPU (HW1) keep wound photos in-house and suit prototypes and studies.
- public: A - public hospitals and health authorities need certified products and procurement-compliant contracts; B or C only for research projects.
Path A — Buy (recommended)
- Dedicated wound-imaging products: imito (CH, MDR-certified, EU-relevant), Swift Medical (CA, with EU/UK presence), Tissue Analytics (US, part of Net Health), plus wound modules in care and EHR software. Check: MDR certificate and intended purpose, EU data hosting, calibration method, offline mode, interface to your care/EHR system (HL7 FHIR, vendor APIs), export of photos with audit trail [3][4].
- Choose A when: you need a device claim, validated accuracy and clinical liability covered by a manufacturer.
Path B — Build on cloud APIs
- Generic vision-language models can describe wound photos, but are not a measurement or validated clinical tool. Use them only for non-clinical helpers (photo quality hints, drafting text from structured values) in an EU region.
- Orchestration via FHIR APIs of the EHR (e.g. Observation, Media/DocumentReference); store photos in an EU-hosted medical-grade store with access logging.
- Building your own measurement function and putting it into clinical use makes you a manufacturer under MDR.
Path C — Self-host (open source)
- Research/prototyping stack: open segmentation models (e.g. nnU-Net, Segment Anything family fine-tuned) with PyTorch; public datasets such as those from wound-segmentation challenges exist but are small and skewed in skin tone and wound type.
- Mobile inference (Core ML, TensorFlow Lite, ONNX Runtime) on phones: HW1 suffices. Intended for research, not for clinical use without conformity assessment.
Path D — Custom training
Worthwhile for manufacturers or hospital groups building a device: needs thousands of expert-annotated images across skin tones, lighting, wound types and devices, multi-rater consensus labels, external validation sites, and a quality management system (ISO 13485, IEC 62304).
State of the art (2026)
- Smartphone segmentation reaches expert-level agreement on well-defined wounds under good conditions in published studies; tissue classification is harder than outline segmentation (slough versus granulation, dark skin, wet or reflective surfaces, dressings residue).
- Depth sensing (LiDAR/stereo) enables wound volume estimates; accuracy depends on cavity shape and capture technique.
- Vendors move from measurement to trajectory prediction (likelihood of healing by week 4 or 12) and population dashboards for care providers; clinical evidence for outcomes is still thinner than for measurement.
- Limits and failure modes: skin-tone bias, poor lighting, curved body parts (heel, sacrum), undermining or tunnelling not visible in photos, infection signs (warmth, odour, pain) invisible to a camera, and automation bias when staff accept outlines unchecked.
Shared resources & combinations
- Same stack (segmentation + calibration): VIS-026 generic measurement and VIS-023 damage assessment from photos share segmentation and scale-calibration methods, but without the medical claim.
- Combines with: HLTH-025 (voice care documentation), HLTH-018 (record summarisation), HLTH-007 (remote monitoring), HLTH-011 (reports), HLTH-006 (skin lesion screening uses classification, not longitudinal measurement).
- Shared mobile edge runtime and FHIR integration layer with other care apps.
Effort, cost & KPIs
- Path A: weeks to a few months for rollout incl. training, devices and EHR/care software integration; pricing typically per user or per site (request quotes).
- Path B/C/D: many months to years for a certified device; QMS and clinical evaluation dominate cost.
- KPIs: agreement with expert measurement, inter-rater variability, documentation time per wound, share of wounds with complete structured documentation, time to escalation of stagnating wounds, healing rate at 12 weeks, photo-quality rejection rate.
People & change
Involve the wound-care lead or head nurse (department-lead) and the nurses who photograph wounds from the start; add IT, the DPO and legal/compliance before the pilot. Wound photos are health data (Art. 9 GDPR): plan a DPIA, a processing agreement with the vendor and clear rules against using private phones (GUIDE-005, GUIDE-007). Patients must know that photos are taken and processed.
Works council: depends. If the app only documents wounds, co-determination is weaker; as soon as it records who photographed what and when, or produces usage statistics per person, it is suited to monitor behaviour and §87(1) no. 6 BetrVG applies (BAG: suitability suffices). Inform the council under §90(1) no. 3 in any case; in public hospitals the staff council applies (GUIDE-004).
AI literacy: support staff with short training on photo technique and on the limits of automatic measurement and classification (Art. 4 AI Act; GUIDE-006). Pilot on one ward with a wound-care champion, compare app measurements with manual ones, then expand (GUIDE-009). Concerns include extra photo work and fear of being checked; answer with time savings per round and the rule that the nurse or physician decides on treatment and can overrule every suggestion (GUIDE-008).
Risks, regulation & compliance
- EU AI Act: software for diagnosis or monitoring of wounds is a medical device; under MDR Rule 11 it is usually class IIa or higher, which makes the AI system high-risk under AI Act Art. 6(1) with Annex I (MDR) [5][6].
- MDR / device claim: measurement and tissue classification for clinical decisions require CE marking by the manufacturer; check intended purpose matches your use. Pure photo storage is different from analysis.
- GDPR: health data is special-category data (Art. 9); photos may show faces, tattoos or home interiors. Need legal basis, DPIA, processor agreement, EU hosting, no private phones or messengers for wound photos; retention per medical record rules.
- Care standards: documentation must meet national expert standards and liability expectations; AI output is a proposal reviewed by qualified staff.
- Fairness: validate across Fitzpatrick skin tones and wound types; monitor performance after deployment.
Getting started
- Measure the baseline: how wounds are measured and documented now, photo handling, escalation delays.
- Shortlist 2-3 MDR-certified products and verify certificate, intended purpose and EU hosting.
- Pilot with 10-20 nurses on real wounds; compare app measurements against expert reference.
- Complete a DPIA and replace private-phone photo practices with the managed app.
- Integrate with care/EHR software and track the KPIs above for 3 months before scaling.
Sources
- DNQP — Expertenstandard "Pflege von Menschen mit chronischen Wunden" — https://www.dnqp.de/
- International Working Group on the Diabetic Foot (IWGDF) guidelines — https://iwgdfguidelines.org/
- imito AG — wound documentation product — https://www.imito.io/
- Swift Medical — Swift Skin and Wound — https://swiftmedical.com/
- Regulation (EU) 2017/745 (MDR) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- MDCG 2019-11 — Guidance on qualification and classification of software — https://health.ec.europa.eu/medical-devices-sector/new-regulations/guidance-mdcg-endorsed-documents-and-other-guidance_en
- EPUAP/NPIAP/PPPIA — International pressure ulcer guideline — https://www.internationalguideline.com/