Melanoma Classification: A Computer Vision Project

Melanoma is the deadliest form of skin cancer partly because it is difficult to distinguish from harmless sunspots and thus goes unnoticed until late in its development. My goal is to use machine learning to distinguish benign spots from melanoma more reliably than the untrained eye.

In this project I built a convolutional neural network (CNN) to classify images of skinspots as benign or malignant with 92% accuracy on the testing set. I also built a simple Streamlit interface where users can upload images of their own skinspots to classify and see a Grad-CAM overlay revealing the specific regions of the skinspot that lead to the model's prediction.

Dataset

Kaggle dataset with 13,879 unique images of skinspots split into a training and testing set:

• Train:
  • Benign - 6289 images
  • Malignant - 5590 images
• Test:
  • Benign - 1000 images
  • Malignant - 1000 images

Structure

• train.py - trains a CNN on the dataset and saves model. 30 epochs maximum, stops early if performance plateaus.
• predict.py - contains functions to classify user-uploaded images and generate a Grad-CAM gradient overlay that visualizes the regions of the image most influential in the model's prediction.
• app.py - calls functions from predict.py to display results on the Streamlit frontend
• utils.py - function for generating accuracy/loss graph and confusion matrix

Model Architecture

After extensive experimentation, I found that the following model architecture led to the highest accuracy and lowest loss while minimizing overfitting:

Performance

Limitations

• During training, all images were resized to 100x100 pixels to ensure fast training cycles, however this came at the cost of accuracy. When downsizing, fine details critical in distinguishing a benign spot from melanoma can be lost.
• User-uploaded images must display the skinspot close-up and in high resolution for the model to generate accurate results. This is a difficult job for most cell phones and is better suited for a digital camera.

Sources

• Melanoma dataset: https://www.kaggle.com/datasets/bhaveshmittal/melanoma-cancer-dataset/data
• Grad-CAM implementation adapted from: https://github.com/ismailuddin/gradcam-tensorflow-2/blob/master/notebooks/GradCam.ipynb