Skip to main content

Web-based PoCT for Skin Cancer Detection using ML and Deep Learning

00:02:00:79

Overview

This research paper presents an innovative web-based approach to preliminary skin cancer detection using machine learning and deep learning models. The study addresses the critical challenge of limited dermatological accessibility due to geographic and economic constraints by developing AI-powered self-diagnosis platforms.

Key Research Highlights

Objective

To develop and compare state-of-the-art deep learning architectures for detecting skin cancer types including melanoma, basal cell carcinoma, and squamous cell carcinoma through web-based interfaces.

Models Evaluated

  • YOLOv8: Moderate accuracy, extremely fast (~5-10 ms/image) - ideal for real-time applications
  • ResNet18: Highest accuracy and consistency across conditions, achieved 100% accuracy on acne, dermatofibroma, and vascular lesions
  • InceptionV3: Best performance for pigmented benign keratosis (93% accuracy)
  • EfficientNet B1: Balanced approach with moderate accuracy
  • Vision Transformers (ViT): Lower accuracy but promising for ensemble techniques

Key Findings

ResNet18 emerged as the most accurate and general-purpose model, achieving exceptional performance across diverse skin conditions. The study demonstrates that strategic model deployment based on lesion type can significantly improve diagnostic accuracy:

  • ResNet18 for melanoma and general-purpose detection
  • InceptionV3 for pigmented lesions
  • YOLOv8 for real-time point-of-care testing

Dataset & Methodology

  • Dataset: 5,000 labeled skin lesion images covering multiple cancer types and benign conditions
  • Techniques: Transfer learning, data augmentation, preprocessing standardization
  • Evaluation Metrics: Accuracy, sensitivity, specificity, computational efficiency

Clinical Significance

The research demonstrates AI's potential to revolutionize dermatological diagnostics by:

  • Enabling accessible early-stage screening
  • Reducing pressure on healthcare systems by identifying high-risk cases
  • Improving survival rates through timely detection
  • Bridging the gap between public access and specialist expertise

Future Directions

The team plans to develop an adaptive mobile application featuring:

  • Automatic model selection based on skin condition classification
  • Real-time confidence scoring and probability analysis
  • Integration of dermatologist feedback for continuous learning
  • Teledermatology capabilities

Publication Details

Authors: Vani Kaushik, Devansh Khodaskar, Chirag Singhal, Anju Gupta
Institutions: Ramdeobaba University, Nagpur, India
Focus: Biomedical Engineering, Computer Science, Electronics Engineering

Access the Research


Research Impact

This work significantly advances AI-driven healthcare by demonstrating the feasibility of web-based, machine learning-powered preliminary skin cancer detection. By providing comparative analysis of leading deep learning architectures and their real-world applicability, the research bridges the gap between academic innovation and practical clinical deployment, ultimately improving accessibility to dermatological care worldwide.