Dermatology AI Classification Challenge

Project Overview

This project aims to build an inclusive machine learning model for dermatology, addressing the underperformance of AI tools on darker skin tones due to limited diverse training data. The model classifies 21 different skin conditions across a range of skin tones using a dataset derived from the FitzPatrick17k dataset.

Dataset

The dataset contains approximately 4500 images categorized into 21 skin conditions. It includes metadata such as the FitzPatrick skin tone scale, diagnostic labels, and quality control annotations.

Files:

• images.zip: Contains the dataset images divided into train and test directories.
• train.csv: Metadata for training images, including labels.
• test.csv: Metadata for test images without labels (to be predicted).
• sample_submission.csv: Sample format for submission.

Data Visualization

Class Distribution

Skin Condition by Fitzpatrick Scale

Model Implementation

Dependencies

The project utilizes the following Python libraries:

• numpy, pandas
• tensorflow, keras
• sklearn
• PIL

Preprocessing

• Loads training and test metadata from CSV files.
• Prepares image paths based on labels and md5hash values.
• Encodes labels using LabelEncoder.
• Splits the training set into training and validation subsets.

Model Architecture

• Convolutional Neural Network (CNN) with layers:
  • Convolutional layers with ReLU activation
  • MaxPooling layers for dimensionality reduction
  • Flatten layer for feature extraction
  • Dense layers for classification
  • Dropout layers for regularization
• Optimized using Adam optimizer and early stopping to prevent overfitting.

Training

• Uses train_test_split to create validation sets.
• Model trained with categorical cross-entropy loss and accuracy metrics.

Evaluation

• Models are evaluated using a weighted average F1-score.

Submission Format

Predictions are stored in a CSV file with the following format:

md5hash,label
16d1e6b4143c88cb158a50ea8bc3a595,acne-vulgaris
aceebbdcfd419fa960ebe3933d721550,folliculitis
85bfb7325d93bac71fcbc08ae0a9ba23,dermatomyositis

Fairness Considerations

• Uses AI fairness and explainability tools to assess model biases.
• Evaluates impact on underrepresented populations.
• Implements data augmentation to ensure balanced representation.

How to Run

Install dependencies:
pip install numpy pandas tensorflow scikit-learn pillow

Run the training script:
python train_test_skin.ipynb

Contribution

Name	GitHub Username	Contribution
Hau Phan	@Hphan777	Performed data augmentation, preprocess data, train model
Shan Ji	@shanji361	Added SMOTE with CNN to handle unbalanced data, made second Kaggle submission, created dataset visualizations
Larissa Zhu	@larissazhu123	Performed feature engineering, built CNN model, performed EDA