Back to 2026 Abstracts
Ai-enhanced Dermoscopy For Early Skin Cancer Detection In Primary Care: Insights From A Retrospective Cohort Study
Raffaele Aguglia, MD1, Moshé Assouline, MD
2, Loïc Van Dieren, BSc
3, Yanis Berkane, MD
3, Haizam Oubari, MD
3, Curtis Cetrulo, Jr., MD
1, Laura Bitton, MD
4, Lucie Duverger, MD
5, Theo Chrelias, MD
6, Georgia Kanellopoulou, MD
6, Leonard Knoedler, MD
1, David Smadja, MD PhD
7, Yohann Dabi, MD PhD
8, Alexandre Lellouch, MD PhD
1.
1Cedars Sinai, Los Angeles, CA, USA,
2Maison Abeille, Medical and Surgical Center for Dermatology & Aesthetics,
75016 Paris, France, Paris, France,
3Vascularized Composite Allotransplantation Laboratory, Center for Transplantation Sciences, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA,
4Medipath, Paris, France,
5Centre National de Dermatopathologistes indépendants, Ivry sur Seine, France,
6Maison Abeille, Medical and Surgical Center for Dermatology & Aesthetics, Paris, France,
7Paris City University, INSERM Paris Cardiovascular Research Center, Paris, France,
8Sorbonne University, Department of Obstetrics and Reproductive Medicine, Paris, France.
PURPOSE: Delays in dermatology consultations pose a significant public health challenge. Artificial intelligence (AI)-assisted dermoscopy has demonstrated promising performance in experimental settings, but evidence from real-world practice is limited.
METHODS: We conducted a retrospective study in a primary care setting to assess the diagnostic performance of an AI-based dermoscopy tool. Images of 403 skin lesions were analyzed, with histopathology as the reference standard. The AI tool was evaluated at multiple malignancy thresholds (0.2=intermediate, 0.5=high risk, and 0.9=Youden’s index). Clinical assessment was performed according to the ABCDE (Asymmetry, Border, Color, Diameter, Evolution) criteria. Diagnostic accuracy was expressed as sensitivity, specificity, and area under the curve.
RESULTS: Malignancy was significantly associated with age>50 years (OR=0.078, p<0.001), recent lesion evolution (OR=1.96, p<0.001), and diameter>0.5 mm (OR=1.86 p<0.001). AUC of the AI tool was 0.814 (95% CI 0.765-0.863). At the 0.2 threshold, sensitivity was 92.3% and specificity 40.8% (AUC 66.6%). Increasing the threshold to 0.5 improved correct classification (62.5%) and specificity (55.9%) but reduced sensitivity (81.7%). The highest Youden index (0.49) was observed at a threshold of 0.90, yielding specificity of 83.6% and sensitivity of 65.4%. Integrating AI with ABCDE improved sensitivity (up to 100% for ABCDE≥2) but reduced specificity. AI misclassified eight malignant lesions.
CONCLUSION: In a real-world primary care setting, using the conservative 0.2 malignancy threshold allowed AI-assisted dermoscopy to maximize sensitivity and detect high-risk lesions that the physician alone might miss. Threshold selection strongly influenced diagnostic trade-offs, underscoring the need to refine AI algorithms to enhance specificity and reduce overdiagnosis.
Back to 2026 Abstracts
