AI Skin Disease Detection System: An Intelligent Tele-Dermatology & Skin Screening Framework Combining Dual Vision-Language AI with Automated Safety Controls and Real-Time Risk Triage

Author

Mr. Ayaz Shaikh, Sachin Kumar, Saif Sayyed, Sanmay Gurudeo, Bhushan Nagdeve, Priyanshu Duratkar


Abstract

 Dermatological conditions represent one of the most prevalent global healthcare burdens, affecting over 1.8 billion individuals worldwide. Early detection of malignant skin lesions, particularly cutaneous melanoma and basal cell carcinoma, is paramount to reducing patient mortality. However, access to certified dermatologists remains severely constrained, especially in non-urban and resource-limited regions. In this work, we present SkinScan AI (v4.0), an enterprise-grade, privacy-aware, multi-modal artificial intelligence screening system designed to bridge the gap between patient self-triage and formal clinical dermatology. SkinScan AI incorporates a dual-provider Vision-Language Model (VLM) architecture leveraging Google Gemini 2.0 Flash and OpenAI GPT-4o Mini operating in synchronized consensus and automatic failover modes. To prevent dangerous diagnostic hallucinations and unnecessary API expenditures, the system integrates a deterministic multi-layer safety governance engine. This includes a pre-LLM computer vision quality gate (evaluating image sharpness, lighting, and skin surface area), a clinical red-flag questionnaire scanner (escalating acute symptoms to emergency status), an automated malignancy keyword scanner, prompt injection defense, and client-side SHA-256 hash caching providing sub-10ms response times for duplicate queries. Furthermore, we implement a 16-feature biomarker extraction pipeline and a calibrated Random Forest ensemble baseline trained on the HAM10000 dataset (N=10,015) using strict patient-level group splitting (GroupShuffleSplit on lesion_id) to eliminate data leakage, and evaluate holdout performance on the ISIC 2018 Task 3 test set (N=1,512). Our baseline model achieves a validation macro ROC-AUC of 0.8406 and holdout test macro ROC-AUC of 0.7958. SkinScan AI is coupled with an interactive tele-dermatology printable summary generator and a doctor-in-the-loop priority queue review portal. We detail the system design, empirical performance, safety rules, regulatory compliance framework, and future deployment roadmap.



Keywords

Tele-dermatology, Vision-Language Models, Multi-Modal Artificial Intelligence, Clinical Safety Governance, Malignancy Escalation, Dual-Provider Consensus, Computer Vision Quality Gating, HAM10000 Dataset, ISIC 2018 Evaluation, Medical AI Ethics.



Full Text:

Download Paper PDF


References


  • Esteva et al., 'Dermatologist-level classification of skin cancer with deep neural networks,' Nature, vol. 542, no. 7639, pp. 115–118, Feb. 2017.
  • Tschandl, C. Rosendahl, and H. Kittler, 'The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions,' Scientific Data, vol. 5, p. 180161, Aug. 2018.
  • Codella et al., 'Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (ISIC),' arXiv preprint arXiv:1902.03368, 2019.
  • Kundu et al., 'Evaluation of vision-language foundation models in clinical dermatology: Opportunities and safety risks,' The Lancet Digital Health, vol. 7, no. 3, pp. e142–e153, 2025.
  • Rotemberg et al., 'A patient-centric dataset of images and metadata for identifying skin cancer in underrepresented populations,' Nature Medicine, vol. 27, no. 1, pp. 154–161, Jan. 2021.
  • A. Haenssle et al., 'Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists,' Annals of Oncology, vol. 29, no. 8, pp. 1836–1842, 2018.
  • Jaworek-Korjakowska et al., 'Computer-aided diagnosis of fungal and bacterial skin infections using deep learning,' IEEE Access, vol. 8, pp. 120540–120551, 2020.
  • Groh et al., 'Evaluating deep learning models for dermatological diagnosis across diverse skin tones,' Journal of the American Academy of Dermatology (JAAD), vol. 86, no. 4, pp. 812–820, 2022.
  • S. Food and Drug Administration (FDA), 'Software as a Medical Device (SaMD): Clinical Evaluation - Guidance for Industry,' Dec. 2017.
  • World Health Organization (WHO), 'Ethics and governance of artificial intelligence for health: WHO guidance,' WHO Guidelines Approved by the Guidelines Review Committee, Geneva, 2021.
  • IEEE Standards Association, 'IEEE P2801 Standard for Recommended Practice for the Quality Management of Medical Artificial Intelligence Data,' 2023.
  • Gemini Team, 'Gemini 2.0 Flash: Technical Report on Vision-Language Capabilities and Fast Inference Architecture,' Google DeepMind Technical Report, 2024.
  • OpenAI, 'GPT-4o Vision System Card and Safety Assessment,' OpenAI Technical Documentation, 2024.
  • A. Maron et al., 'Precision of automated melanoma detection: A systematic benchmark of deep learning algorithms,' European Journal of Cancer, vol. 155, pp. 176–184, 2021.
  • Barata et al., 'Deep learning for skin cancer detection with clinical and dermoscopic images,' IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 7, pp. 2704–2712, 2021.
  • Shaikh, S. Kumar, S. Sayyed, S. Gurudeo, B. Nagdeve, and P. Duratkar, 'AI Skin Disease Detection System Technical Specification,' Department  of  Computer  Science  and  Engineering,  Nagpur  Institute  of  Technology,  GitHub  Repository: sachin9326/Ai-skin-disease-detection-system-app, Oct. 2026.

 

 

Share your valuable work from Social Media Buttons