The ACCURACY Study
The ACCURACY Study
Sangers T, Reeder S, van der Vet S, et al. Validation of a Market-Approved Artificial Intelligence Mobile Health App for Skin Cancer Screening: A Prospective Multicentre Diagnostic Accuracy Study. Dermatology 2022;238:649–656.
Objective:To validate the diagnostic accuracy of the SkinVision app, which uses a convolutional neural network (CNN) to detect premalignant and malignant skin lesions
Study Design
Prospective, cross-sectional, multicentre diagnostic accuracy study (2020)
- Conducted at two dermatology outpatient clinics in the Netherlands (Erasmus MC Cancer Institute & Albert Schweitzer Hospital)
- Each lesion was first examined by a clinician, then assessed using the SkinVision app (iOS or Android)
- App risk outcomes were compared to histopathology (when indicated) or dermatologist diagnosis
- Reporting followed Standards for Reporting Diagnostic Accuracy (STARD) 2015 guidelines1
Study Population
Adults ≥18 years with ≥1 suspicious skin lesion
- Excluded: prior biopsy, unclear diagnosis, technical failures, or inability to consent
785 lesions (from 372 patients)
Median age: 71 years
Gender: 50.8% women
Fitzpatrick skin type: >80% type I–II
Primary Outcome:
- Sensitivity and specificity of the SkinVision app in detecting premalignant and malignant lesions.
Key Results
Sensitivity
86.9%
(90% Cl 82.3-90.7)
The SkinVision app correctly identified 239 of the 275 (pre) malignant lesions as high-risk
Specificity
80.1%
(95% Cl 75.7-84.1)
The Specificity of the SkinVision app was set at 80.1% (based on benign control lesions)
Sensitivity by cancer type2:
Malignant melanoma (n=38)
92.1%
Squamous cell carcinoma (n=72)
98.6%
Basal cell carcinoma (n=265)
93.6%
Premalignant lesions (n=131)
90.8%
Positive predictive value (PPV):
61.3%
Negative predictive value (NPV):
90.9%
Overall results:
785 total lesions assessed:
418 suspicious (53.2%), 367 benign (46.8%)
Confirmed (pre)malignant:
275 lesions(35%)
Overall sensitivity:
86.9% (95% CI: 82.3%–90.7%)
Correctly identified:
239/275 (pre)malignant lesions as high-risk
Sensitivity
86.9%
(90% Cl 82.3-90.7)
The SkinVision app correctly identified 239 of the 275 (pre) malignant lesions as high-risk
Specificity
80.1%
(95% Cl 75.7-84.1)
The Specificity of the SkinVision app was set at 80.1% (based on benign control lesions)
Sensitivity by cancer type2:
Malignant melanoma (n=38)
92.1%
Squamous cell carcinoma (n=72)
98.6%
Basal cell carcinoma (n=265)
93.6%
Premalignant lesions (n=131)
90.8%
Positive predictive value (PPV):
61.3%
Negative predictive value (NPV):
90.9%
Overall results:
785 total lesions assessed:
418 suspicious (53.2%), 367 benign (46.8%)
Confirmed (pre)malignant:
275 lesions(35%)
Overall sensitivity:
86.9% (95% CI: 82.3%–90.7%)
Correctly identified:
239/275 (pre)malignant lesions as high-risk
Key Takeaways
- The SkinVision app achieved high sensitivity (86.9%) and specificity (80.1%) for detecting skin (pre)malignancies in a clinical setting – performance that is comparable to experienced dermatologists (sensitivity 76.9%, specificity 89.1%)3.
- These findings support the app’s potential as a reliable self-assessment and triage tool, while also highlighting the need for further validation in lay users and across diverse skin types.