Skin Cancer Classification

To classify skin cancer images as benign or malignant.

Dataset

The raw dataset is provided under raw_dataset/ to be used as the main data source.

Dataset directory structure

raw_dataset/
|___ test/
|    |___ Benign/
|    |    |___ 6299.jpg
|    |    |___ 6300.jpg
|    |    |___ ...
|    |___ Malignant/
|         |___ 5602.jpg
|         |___ 5603.jpg
|         |___ ...
|
|___ train/
|    |___ Benign/
|    |    |___ 1.jpg
|    |    |___ 2.jpg
|    |    |___ ...
|    |___ Malignant/
|         |___ 1.jpg
|         |___ 2.jpg
|         |___ ...

Model

The model used in here is YOLOv8 (small) classification model

Directory Structure

Place all the files in same directory as the following:

├─── raw_dataset/   raw data source
├─── models/        contains classification models 
├─── utils/         contains helper functions and classes
├─── demo-classification.ipynb  jupyter notebook demonstrate the classification 
├─── demo-preprocessing.ipynb  jupyter notebook run the preprocessing demo 
├─── classifier.py      contains SkinCancerClassifier class 
└─── clean_dataset.py  command line tool to create clean dataset for training

Installation

Clone

git clone https://github.com/mshenoda/skin-cancer-classification.git

Create environment

cd diffugen
conda create -n skin python=3.12
conda activate skin

Install requirements

PyTorch with CUDA Dependency

pip install torch==2.3.1 torchvision==0.18.1 --index-url https://download.pytorch.org/whl/cu121

Other packages

pip install -r requirements.txt

Usage

Runing demo

Recommended to use VSCode for running demo notebooks

• preprocessing techniques:
  • run demo-preprocessing.ipynb
• classification / postprocessing evaluation:
  • run demo-preprocessing.ipynb

Data cleaning

python clean_dataset.py raw_dataset clean_dataset

Training

cd models\yolov8s-cls
python train.py

Training results are under models\yolov8s-cls\runs\classify\

Examples - Data Cleaning:

Examples - Prediction:

Citation

If you find this useful for your research, please cite:

@inproceedings{skin-cancer-classification,
  title     = {Skin Cancer Classification},
  author    = {Shenoda, Michael}
  year      = {2024}
}