Training a CNN From Scratch For Melanoma Detection Using Keras

Dr. Abder-Rahman Ali
2 min readDec 29, 2019


I will be sharing a script using Keras for training a Convolutional Neural Network (CNN) from scratch for melanoma detection. You can view the code from this GitHub repository.

Before proceeding however, make sure that you structure the data as follows (the numbers represent the number of images in each file):

You can download the data from, here. I used two classes as you can see from the figure above (nevus and melanoma). For training, I kept 374 images in each class to keep the data balanced.

The results will not be optimal, as the purpose is to show how one can train a CNN from scratch.

What variables to edit in the code?

You need to edit the following variables to point to your data:

train_directory (path to your training directory)

validation_directory (path to your training directory)

test_directory (path to your testing directory)

What should you expect (outputs)?

Training and validation accuracy

Training and validation loss

Test accuracy

This will be a value. In my case, the test accuracy was around 77.1%.

ROC curve

In addition to some other values (i.e. confusion matrix) that will be displayed on the console.

The next post shows how to train a CNN from scratch for melanoma detection, but with data augmentation. You can find the next post, here.



Dr. Abder-Rahman Ali

Research Fellow @ Massachusetts General Hospital/Harvard Medical School |