Skin Lesion Classification (DermaMNIST)

Final project for CS-UH 2220: Machine Learning

---

Overview

This project performs multi-class classification of dermatological images using the DermaMNIST dataset.

The repository contains:

• Exploratory Data Analysis (EDA)
• Deep learning model training using ResNet-18 (transfer learning)

---

Dataset

• 10,015 RGB images
• 7 skin lesion classes
• Significant class imbalance
• Source: MedMNIST (DermaMNIST)

---

Repository Structure

skin-lesion-classification/
│
├── notebooks/
│   ├── 01_data_exploration.ipynb
│   └── 02_resnet18_training.ipynb
│
├── figures/
│   ├── class_distribution.png
│   ├── sample_images.png
│   ├── pixel_histogram.png
│   └── confusion_matrix.png
│
├── README.md
├── requirements.txt
└── .gitignore

---

Model

• ResNet-18 (ImageNet pretrained)
• CrossEntropyLoss
• Cosine Learning Rate Scheduler
• Data augmentation
• Evaluation using precision, recall, and F1-score

Final Performance

• Test Accuracy: ~80%
• Weighted F1-score: ~0.79
• Macro F1-score: ~0.63
• Melanoma recall monitored separately

---

How to Run

Recommended: Google Colab

1. Run 01_data_exploration.ipynb
2. Run 02_resnet18_training.ipynb (GPU recommended)

Dependencies are listed in requirements.txt.

---

Key Takeaways

• Severe class imbalance impacts minority class performance.
• Transfer learning significantly improves generalization.
• Monitoring per-class metrics is crucial in medical ML tasks.