Saltar al contenido principal

Limitations

This page documents the known limitations of automated VASI scoring, so that sponsors and CROs can take them into account during protocol design.

Depigmentation is measured, not diagnosed​

The model measures the extent of light skin, whatever its cause. It does not distinguish vitiligo from other conditions that also cause light skin lesions, such as pityriasis versicolor, pityriasis alba, post-inflammatory hypopigmentation or idiopathic guttate hypomelanosis.

Mitigation: Diagnosis is the investigator's responsibility and is confirmed at enrolment. The protocol should specify how to handle a patient who develops a second condition with light skin lesions during the study.

Depigmentation and hypopigmentation​

The two words are not synonyms. Depigmented skin has lost its pigment completely and looks chalk-white; hypopigmented skin keeps some pigment and looks lighter than the skin around it. Vitiligo lesions are depigmented, and hypopigmented skin appears with them: at the edge of a lesion in trichrome vitiligo, in new lesions, which pass through a hypopigmented phase, and in lesions that are repigmenting.

The model reports both as one class, because a photograph under visible light does not show reliably where one ends and the other begins. This site uses "depigmentation" for that class, which is the word of the VASI.

Degree of depigmentation​

Manual VASI multiplies the extent of each area of vitiligo by its degree of depigmentation, graded in seven steps from 100% (no pigment) down to 10% (specks of depigmentation). Their product is the surface of the area that has lost its pigment, which is what the automated score measures by counting depigmented pixels, so it does not grade the degree separately; see how the automated score relates to the VASI.

The images the model learned from were annotated as precisely as each image allowed, leaving out the islands of pigment inside repigmenting lesions, and the model does the same. The smallest dots of repigmentation around the hair follicles may still go undetected, depending on the distance and on how large the area is in the photograph. The skin they cover then counts as depigmented, so early repigmentation shows later in the automated score than to a rater. The model also counts hypopigmented skin, which keeps some pigment, in full.

The automated and the manual score measure the same quantity, and they can still differ in three more ways:

  • Rounding: The rater rounds the degree to seven steps and the extent to tenths of a hand unit; the automated score is continuous.
  • The unit of surface: The rater's unit is the patient's own hand, which stands for about 1% of the body; the shares of the views are fixed adult values. The two units agree on average, not for every patient.
  • Curved surfaces: A photograph flattens the body, so skin that turns away from the camera, at the sides of the trunk and the limbs, covers fewer pixels than its surface, and a lesion there counts for less. Photographing each side of a site from straight in front of it keeps this small.

Mitigation: Each view is photographed at the shortest distance at which it fits in the frame, the same at every visit, as the imaging protocol sets out, which keeps the dots as large as possible in the photograph. Until the agreement with the rater's VASI is reported on repigmenting lesions, a sponsor whose primary endpoint is a VASI response can run a fit-for-purpose validation on its own study population.

Very fair skin​

Contrast between pigmented and depigmented skin is the main visual cue for the model. On the lightest skin types (Fitzpatrick I and II) that contrast is low, which is the same reason clinicians use a Wood's lamp to examine these patients.

The model's agreement with the dermatologists' reference on these skin types is similar to the other groups and meets its acceptance criterion; see Performance across skin phototypes.

Mitigation: If the study population is predominantly of the lightest skin types, this should be discussed during protocol design.

Ultraviolet (Wood's lamp) images​

Photographs taken under a Wood's lamp or any other ultraviolet light are not supported; the AI scores photographs taken under visible light. Some trial protocols have the rater score the VASI under a Wood's lamp, which shows depigmentation on fair skin that visible light hides, so on those patients the rater's VASI can be higher than the automated one.

Mitigation: The rater keeps the Wood's lamp for the clinical examination, and the photographs for the AI are taken under visible light. Studies that need ultraviolet imaging should discuss it during protocol design.

Regions outside the capture set​

The score covers only the skin that the standard views show. Regions that are not photographed, or that are hidden by the camera angle or by the patient's position, are not measured.

The full capture set covers 95.5% of the body surface. It leaves out the back of the head and the scalp, which hair covers (3.5%), and the genitals, which are never photographed (1%). The VASI counts the genitals, usually within the trunk, and some protocols also score them on their own, which stays with the rater. The ears are outside the facial area of the F-VASI and are not covered by any view. The T-VASI therefore reads slightly low in a patient whose vitiligo involves these areas.

The face is scored on two views turned about 45 degrees, one for each half. Skin near the midline or right in front of the ear still turns partly away from the camera in its view, so a lesion there counts for slightly less than its surface.

Mitigation: The imaging protocol can add views for regions the study needs, during protocol design.

Endpoints not measured​

The platform measures the extent of depigmentation. It does not measure:

  • Disease activity signs: confetti-like depigmentation, trichrome lesions and the Koebner phenomenon
  • Vitiligo Disease Activity (VIDA) score, which is based on the patient's history rather than on the image
  • Clinician and patient global impressions, such as the CGIS and the PGIS
  • Patient-reported outcomes, such as the Vitiligo Noticeability Scale (VNS)

Mitigation: These endpoints remain with the investigator or the patient and are collected in the EDC alongside the automated VASI.

Photograph-based assessment​

The AI analyses photographs of the skin, not the patient. What only an in-person examination shows, such as leukotrichia, the response to a Wood's lamp or the patient's history of new lesions, cannot be read from an image.

Mitigation: The imaging protocol standardises the views, the distance and the lighting, and the DIQA quality gate rejects images that do not meet minimum quality standards for focus, lighting, framing and resolution. The investigator reviews every report alongside the clinical examination.

Agreement with the investigator's VASI​

The segmentation model has been validated against the consensus of several dermatologists on individual photographs. The agreement of the total automated VASI with the VASI assigned by dermatologists on the same patients has not yet been reported; see Clinical evidence.

Mitigation: Until that agreement is reported, sponsors can run a fit-for-purpose validation on their own study population during study setup.

Reference standard​

The reference the model is measured against is the consensus of several dermatologists who traced the boundaries of depigmentation on each photograph. It is the best available approximation of the true extent, but it is not an objective measurement, and the model cannot be more accurate than the reference it was trained and tested against.

Mitigation: Using the consensus of several annotators reduces the bias of any single one.

Model version specificity​

All performance metrics reported in this documentation apply to a specific validated model version. Model updates, including retraining, architecture changes or threshold adjustments, require full re-validation under IEC 62304 before deployment.

Mitigation: The model version is locked at study initiation. No mid-study model updates occur, so every patient in a trial is scored by the same model and endpoint integrity is preserved throughout the study.

Decision support, not autonomous diagnosis​

The system provides severity scoring to support clinical decisions. It does not replace clinical judgement, and it does not make autonomous diagnostic or treatment decisions. All AI-generated scores should be interpreted by qualified healthcare professionals within the context of the patient's overall clinical presentation.

How limitations are managed​

All limitations documented on this page are tracked within the formal risk management process (ISO 14971) and the software development lifecycle (IEC 62304). Each limitation has been assessed for clinical risk, and mitigations have been implemented where the residual risk is not already acceptable.

The post-market clinical follow-up (PMCF) programme under MDR Annex XIV continuously monitors real-world performance. Any new limitation identified through post-market surveillance triggers a formal risk assessment and, if necessary, corrective action.