Extreme Learning Machine-Mixer: An Alternative to Multilayer Perceptron-Mixer and Its Application in Skin Cancer Detection Based on Dermoscopy Images

Abstract

Skin cancer is a serious disease caused by abnormal skin cell growth, often due to excessive UV exposure. Dermoscopy helps physicians examine skin lesions, aiding in the detection of benign and malignant growths not easily seen with the naked eye. Recently, a great interest has been in developing computer-aided diagnostics that can diagnose skin cancer using artificial intelligence (AI). In this paper, we develop a novel AI approach, Extreme Machine Learning (ELM)-Mixer, and apply it to skin cancer detection based on dermoscopy images. The uniqueness of the ELM-Mixer is demonstrated through faster training times and reduced computational complexity compared to deeper models, owing to its single-pass learning approach. We develop ELM-Mixer architecture as an alternative to Multilayer Perceptron (MLP)-Mixer. As ELM is a training-free approach, we adopt this feature to form an ELM-Mixer where the computation burden of the MLP-Mixer will be alleviated. ELM-Mixer architecture begins with the patch embedding part and a mixer layer comes after the patch embedding part. In the mixer layer, layer normalization, token, and channel mixing operations are used. After the mixer layer, an ELM head layer is located for calculating the output of the network with classical Moore-Penrose matrix inverse operation. The HAM10000 skin cancer dataset is considered in the experimental works and various performance evaluation metrics are used. The experiments show that the proposed model produced a better overall 94.7% precision, 94.6% recall, 94.6% accuracy, 99.1% specificity, and 94.6% F1-score than the compared state-of-the-art achievements.