Revolutionizing
Skin Cancer
Diagnosis

Precision Skin Lesion Classification Using Deep Learning

Advanced Deep Learning Research Project

University of Toronto APS360

Project Overview

🎯

Purpose

Design a deep learning model to classify various types of skin lesions, supporting dermatologists with AI-powered diagnostic assistance for improved accuracy and accessibility.

💡

Motivation

• 80,000 annual skin cancer cases in Canada

• 99% five-year survival rate with early detection

• Only ~60% accuracy in visual examination

• Limited dermatologist accessibility

🚀

Vision

A foundational step towards deployable applications that integrate seamlessly into clinical workflows, enhancing diagnostic efficiency and supporting healthcare professionals.

Meet the Team

The minds behind SkinAI Classifier

Team Member 1

Vishwas Puri

Developer

 Team Member 2

Rayan Ahsan

Developer

 Team Member 3

Adyan Hossain

Developer

 Team Member 4

Kashan Ahmad

Developer

Technical Deep Dive

ResNet-18 Architecture

Input Layer
→
Conv Blocks
→
Skip Connections
→
FC Layer (7 outputs)

Pretrained ResNet-18 with reconfigured final fully connected layer for 7-class skin lesion classification

Classification Pipeline

📷

Input Images

Dermoscopic images

⚙️

Preprocessing

224x224 resize, normalization

🔄

Data Augmentation

Address class imbalance

🧠

ResNet-18 Model

Deep learning classification

📊

Prediction

7-class output

HAM10000 Dataset

Total Images: 10,015

Seven Diagnostic Categories:

  • • Melanocytic nevi (nv)
  • • Melanoma (mel)
  • • Benign keratosis-like lesions (bkl)
  • • Basal cell carcinoma (bcc)
  • • Actinic keratoses (akiec)
  • • Vascular lesions (vasc)
  • • Dermatofibroma (df)

Data Processing

Data Cleaning:

  • • Removed corrupt/duplicate images
  • • Handled null values
  • • Resized to 224x224 pixels
  • • Applied normalization

Challenges Addressed:

  • • Class imbalance via augmentation
  • • Computational optimization
  • • Training time efficiency

Live Demo

Experience our skin cancer classification model in action. Upload an image and see real-time AI-powered diagnosis.

Model Performance

✓
85.2%
Training Accuracy
📊
78.9%
Validation Accuracy
⚡
0.45
Final Loss
🎯
7
Classification Classes

Performance Analysis

Model Strengths

  • ✓ Strong performance on common lesion types
  • ✓ Effective transfer learning from ImageNet
  • ✓ Robust feature extraction with ResNet-18
  • ✓ Efficient computational requirements

Areas for Improvement

  • → Class imbalance in rare lesion types
  • → Enhanced data augmentation strategies
  • → Ensemble methods for improved accuracy
  • → Clinical validation requirements

Reports

View our project proposal and progress report.

Select a report to view the PDF.

Research Impact

APS360: Applied Fundamentals of Deep Learning
University of Toronto

Advancing Medical AI

This project represents a significant step forward in applying deep learning to medical diagnosis, demonstrating the potential for AI to support healthcare professionals in critical decision-making processes. Our work contributes to the growing field of medical AI and showcases the practical application of convolutional neural networks in real-world healthcare scenarios.

10,015
Images Processed
7
Lesion Types Classified
85.2%
Peak Accuracy Achieved