skin-lesion-analysis

This repository contains code to classify images of skin lesions to detect melanoma.

Data

Get data from kaggle:

kaggle datasets download -d wanderdust/skin-lesion-analysis-toward-melanoma-detection

Unzip to directory skin-lesions/

Training images are drawn uniformly from the three classes (Melanom, Nevus and Seborrheic Keratosis), cropped to squares and rescaled to 224 by 224 RGB pixels. Images are read and transformed during training, use multiple threads to make training faster.

Models

Resnet

The script detect_melanom.jl finetunes a pretrained ResNet (He et al. 2016) with depth 50 to classify between Melanom, Nevus and Seborrheic Keratosis. Accuracy approaches ~65%.

Vision Transformer

Setting the model to transformer finetunes a pretrained vision transformer. It takes longer to train, but the achieved accuracy is similar to the resnet.