YOLOv8-Based Deep Learning Approach for Real-Time Skin Lesion Classification Using the HAM10000 Dataset

This repository contains code for the paper titled YOLOv8-Based Deep Learning Approach for Real-Time Skin Lesion Classification Using the HAM10000 Dataset. The ISIC2018 Task3 dataset (HAM10000) is used for this paper.

This paper was presented at the 2024 International Conference on e-health Networking, Applications and Services (HealthCom) at Nara, Japan.

DOI: https://doi.org/10.1109/HealthCom60970.2024.10880715

Dataset

The project uses the ISIC2018 Task3 dataset (HAM10000), which contains skin lesion images classified into 7 different categories:

• Melanoma
• Melanocytic nevus
• Basal cell carcinoma
• Actinic keratosis / Bowen's disease (intraepithelial carcinoma)
• Benign keratosis (solar lentigo / seborrheic keratosis / lichen planus-like keratosis)
• Dermatofibroma
• Vascular lesion

Project Structure

├── data_processing.py       # Script for processing ISIC2018 dataset
├── train.py                 # Script for training YOLOv8 models
├── test.py                  # Script for evaluating trained models
├── train_with_aug.log       # Training log with data augmentation
├── train_without_aug.log    # Training log without data augmentation
├── result_with_aug/         # Results directory for models trained with augmentation
├── result_without_aug/      # Results directory for models trained without augmentation
└── runs/                    # Directory containing visualization results

Installation

1. Clone this repository:

https://github.com/uts58/yolov8-ham10000.git
cd yolov8-ham10000

2. Install the required packages:

pip install ultralytics scikit-learn matplotlib seaborn pandas

Usage

Data Preparation

The data_processing.py script organizes the ISIC2018 dataset into the appropriate directory structure:

python data_processing.py

Training

To train YOLOv8 models on the skin cancer dataset:

python train.py

This script trains five different YOLOv8 model sizes (nano, small, medium, large, and extra large) on the dataset for 30 epochs.

Evaluation

To evaluate the trained models:

python test.py

This script generates classification reports and confusion matrices for each model.

Models

The project uses the following YOLOv8 classification models:

• YOLOv8n (nano): 1.4M parameters
• YOLOv8s (small): ~6M parameters
• YOLOv8m (medium): ~17M parameters
• YOLOv8l (large): ~37M parameters
• YOLOv8x (extra large): ~68M parameters

Results

The results are saved in the following directories:

• result_with_aug/: Contains models trained with data augmentation
• result_without_aug/: Contains models trained without data augmentation

For each model, the following files are generated:

• Classification report (CSV)
• Confusion matrix (PNG)

Citation

Cite our paper by

@INPROCEEDINGS{10880715,
  author={Saha, Utsha and Ahamed, Imtiaj Uddin and Imran, Md Ashique and Ahamed, Imam Uddin and Hossain, Al-Amin and Gupta, Ucchash Das},
  booktitle={2024 IEEE International Conference on E-health Networking, Application & Services (HealthCom)}, 
  title={YOLOv8-Based Deep Learning Approach for Real-Time Skin Lesion Classification Using the HAM10000 Dataset}, 
  year={2024},
  volume={},
  number={},
  pages={1-4},
  keywords={Measurement;Deep learning;Accuracy;Computational modeling;Skin;Real-time systems;Lesions;Ensemble learning;Skin cancer;Diseases;Skin Cancer;YOLOv8-Classification;Skin Lesion Classification;HAM10000 Dataset;Transfer Learning;Deep Learning},
  doi={10.1109/HealthCom60970.2024.10880715}}

Acknowledgments

• Ultralytics YOLOv8
• ISIC2018 Challenge