Melanoma is an extremely dangerous type of skin cancer most commonly caused by exposure to UV light. The highest incidence rate of Melanoma is in Australia and New Zealand, in addition to being the fourth most common cancer diagnosed in New Zealand. Although Melanoma can be fatal, if it is discovered quickly it can be cured relatively easily. Hence there is growing motivation to develop methods of detecting Melanoma early.

One method of detecting Melanoma is through the usage of machine learning on medical images. The aim of this project was to develop a suitable machine learning model that could accurately classify images as melanoma or non-melanoma based on the contents of the images.

The image set we used included approximately 4000 images of skin images, which were split into two main categories: Melanoma and Non-Melanoma. The Melanoma images were further split into two sub categories: Close Up (which are photos taken with cameras) and Dermoscopy (photos taken using a device that is similar to a microscope), The Non-Melanoma images were also split into Close Up and Dermoscopy but also had 3 further sub categories being BCC, SCC (two other types of skin cancer) and Other non skin cancer images.

The resulting algorithm was tested against a range of images types and qualities, with promising results as well as some key challenges and lessons learned, including:

• Varying image quality, lighting, and inclusion of non-relevant content such as rulers and hairs impacting accuracy of algorithm.
• Background clutter in images such as occlusion impacting the model’s ability to find the subject of the image