# Automated Machine Learning (AutoML) for the Diagnosis of Melanoma Skin Lesions From Consumer-Grade Camera Photos

## TL;DR

This study shows that a no-code AI tool can accurately diagnose melanoma from skin lesion photos, offering potential for non-experts to develop clinical AI tools.

## Contribution

Demonstrates the use of a no-code AutoML platform for melanoma diagnosis with performance comparable to expert-developed models.

## Key findings

- The AutoML algorithm achieved 84.4% accuracy in classifying melanoma and non-melanoma lesions.
- It correctly identified 83.3% of melanoma cases and 85.7% of non-melanoma cases in the test set.

## Abstract

Background: In recent years, there has been much speculation about the role of artificial intelligence (AI) and machine learning in dermatology. Advances in computer vision have increased the potential for automated diagnosis of images. However, there remains a gap between the technological development of the algorithms and their real-world implementation. This study aims to develop and test an automated machine learning (AutoML) algorithm for the diagnosis of melanoma, with no technical or coding skills required by the operator.

Methods: The Skin Cancer Detection Dataset from the University of Waterloo Vision and Image Processing Lab contains 206 images sourced from the public databases DermIS and DermQuest. The dataset was split into two groups: training data (n=174) and testing data (n=32). A machine learning algorithm was created using ‘Teachable Machine’, trained on the training data, to differentiate between melanoma and non-melanoma skin lesions.

Results: The AutoML algorithm identified 12/14 non-melanoma images and 15/18 melanoma images in the testing dataset. The overall accuracy was 84.4%, with a sensitivity of 83.3% and a specificity of 85.7%.

Conclusions: Existing literature has tested a range of different machine learning algorithms on the same dataset. These have often required expertise in machine learning and the ability to code. The results of this study, using a no-code tool, perform comparably to existing efforts and suggest that there is potential for future clinical AI algorithms to be developed by doctors even without any technical expertise as long as they have access to relevant local data.