Development of an AI-driven dermatoscopic image analysis platform for early detection and classification of pigmented skin lesions in resource-limited settings

Project Abstract

This study presents the development and evaluation of an AI-driven dermatoscopic image analysis platform designed for early detection and classification of pigmented skin lesions in resource-limited settings. The platform integrates on-device deep learning models with a lightweight cloud-assisted workflow to enable rapid, affordable, and accessible dermatologic assessment, particularly in underserved communities with limited access to specialist care and diagnostic infrastructure. We collected a diverse dataset of dermatoscopic images from collaborating clinics across three low- to middle-income regions, annotated by board-certified dermatologists to establish robust ground truth labels for common pigmented lesions, including malignant melanomas, dysplastic nevi, and benign nevi. Data preprocessing included standardization of illumination, color normalization, and lesion segmentation to ensure model resilience against variability in imaging devices and acquisition conditions. Our approach leverages a multi-branch convolutional neural network augmented with attention mechanisms to capture both global lesion morphology and local dermoscopic features such as pigment networks, streaks, globules, and regressive patterns. To address data scarcity and class imbalance, we employed transfer learning from pre-trained skin cancer models, synthetic data augmentation, and stratified sampling to maintain representative performance across lesion categories and skin phototypes. The platform supports offline operation on low-cost mobile devices, with a modular architecture that enables optional cloud synchronization for expert consultation and longitudinal tracking. We implemented an interpretable AI framework that generates saliency maps and lesion attribute explanations to enhance clinician trust and patient communication while adhering to medical device regulatory considerations. Performance was evaluated using standard metrics (accuracy, sensitivity, specificity, area under the ROC curve) on an held-out test set, with stratified analyses by skin type and acquisition device. Cross-validation results demonstrated high sensitivity for malignant lesions and robust specificity for benign conditions, outperforming baseline traditional computer vision approaches. A prospective pilot study in three clinics assessed real-time usability, turnaround time, and impact on diagnostic decision-making by non-specialist healthcare workers, revealing improvements in triage efficiency and referral accuracy. The platform also incorporated risk-stratified decision support, providing confidence scores and recommended next steps aligned with established dermatology guidelines. Privacy and data governance were prioritized through on-device inference, encrypted data transmission, and adherence to ethical approvals and consent protocols. Limitations include potential biases due to geographic clustering, variability in image quality, and the need for ongoing model fine-tuning to accommodate novel lesion types and emerging dermoscopic patterns. Future work will expand the dataset to additional populations, integrate multimodal inputs such as clinical photographs and patient history, and explore federated learning to enhance generalizability while preserving data privacy. The study demonstrates that an AI-assisted dermatoscopic platform can empower frontline clinicians in resource-constrained settings to perform accurate, timely skin cancer risk assessment, enabling earlier interventions and improved patient outcomes with scalable, low-cost deployment.

Project Overview

What This Project Is About

---

This project explores a computer-based tool that helps doctors and health workers identify and classify pigmented skin lesions using simple images of the skin. It combines easy image analysis with clear guidance to determine whether a lesion is likely benign or potentially harmful, focusing on settings with limited medical resources.

---

---

The Problem It Addresses

---

---

---

Objectives of the Project

---

1. Develop a simple dermatoscopic image analysis tool that works on common devices.
2. Teach the system to distinguish between common benign lesions and suspicious ones.
3. Make the tool easy to use for non-specialists with minimal training.
4. Evaluate accuracy, speed, and reliability in real-world settings.
5. Assess how the tool could fit into existing health workflows.

---

---

What You Will Do Step by Step

---

1) Collect a diverse set of labeled skin lesion images (with consent). 2) Build a simple image-processing pipeline that extracts helpful features. 3) Train a basic machine-learning model to classify lesions. 4) Test the model’s accuracy on separate images. 5) Create a user-friendly interface for frontline health workers. 6) Run small field trials to understand usability. 7) Compare performance with standard guidelines. 8) Document limitations and potential improvements.

---

---

Expected Outcome

There will be a working, easy-to-use tool that can screen pigmented skin lesions from photos, with documented accuracy and guidance for when to seek further care. It aims to improve early detection, supporting healthcare workers in resource-limited settings.