Skin Disease Diagnosis Using Computer Vision

Which supervised learning model (e.g., CNN, ResNet, Vision Transformer, or EfficientNet) provides the most accurate classification of skin lesions using dermatoscopic images, and how does including patient metadata (such as age, sex, and lesion location) affect model performance?

Problem Statement

Describe the motivation for this project, why it is relevant, and what its impacts are.

EXAMPLE: Given the substantial daily output of responses, the identification and mitigation of ChatGPT's biases become critical, safeguarding both the multitude of users and the far-reaching consequences they may influence.

Key Results

fter comparing CNN, ResNet, and EfficientNetB0, we found that: EfficientNetB0 achieved the best overall performance across accuracy, recall, and F1-score, particularly in detecting critical classes like melanoma. Including patient metadata (age, sex, anatomical site) improved performance modestly This confirms that: The model architecture matters significantly in lesion classification. Contextual data (metadata) enhances accuracy, especially for clinically important classes. AI’s broader impact: This research shows that well-designed AI can support medical diagnosis, but: Model bias (e.g., by skin tone or age group) needs mitigation. Ethical deployment requires transparency and clinician oversight.

Methodologies

Target Lesion Types Top 4 Diagnoses: Melanocytic nevi (NV) (Most common benign) Melanoma (MEL) (High-risk malignant) Basal cell carcinoma (BCC) (Common skin cancer) Benign keratosis-like lesions (BKL) (Non-cancerous)

*To accomplish this, we created the models using CNN, ResNet, Vision Transformer, to undergo surpervised machine learning to classify dermatoscopic images into the top 4 diagnosis. Each data class was evenly training, validating, and testing stages.

Data Sources

ISIC 2019 Challenge Dataset 25,331 high-quality dermatoscopic images of skin lesions Each image is labeled with one of nine diagnostic categories, including:

• Melanoma (MEL)
• Melanocytic nevus (NV)
• Basal cell carcinoma (BCC)
• Actinic keratosis (AK)
• Benign keratosis-like lesions (BKL)
• Dermatofibroma (DF)
• Vascular lesions (VASC)
• Squamous cell carcinoma (SCC)
• Unknown (UNK)

Metadata file is included for each image, containing: -Patient’s approximate age (age_approx) -Sex (sex) -Anatomical site of the lesion (anatom_site_general) -Lesion ID (can be used to track images from the same lesion)

Kaggle Dataset: Link to Kaggle Dataset*

Technologies Used

• Python
• pandas
• OpenAI API
• CNN
• ResNet
• Vision Transformer

Authors

This project was completed in collaboration with:

• Alisa Teige (john.doe@example.com)
• Grish Tautam (john.doe@example.com)
• Emmanuel Moses (john.doe@example.com)
• Tory Leone (john.doe@example.com)
• Darelle Herrera (jane.smith@example.com)