Predicting non-melanoma skin cancer via a multi-parameterized artificial neural network

Abstract

A significant and important part of our life is our health, and integrating computers into our daily lives has the potential to improve our health. Healthcare could undergo a radical change if a machine learning model that can identify a benign or cancerous skin mole from a picture is developed. This paper performs a comprehensive review with an emphasis on the developing body of research on the use of convolutional neural networks (CNNs) to classify skin lesions. The focus is on classifiers that are especially made for skin lesions; techniques that use CNNs only for segmentation or classification of thermoscopic patterns are not included. The paper explores the difficulties in comparing various processes and discusses the issues that need to be resolved in other studies. The implementation uses the Python programming language and a CNN model in Keras with TensorFlow as the backend. ISIC databases that are open to the public are used for training and testing in this research, which is carried out on the Kaggle platform. The collection consists of balanced photos of skin moles that are either benign or cancerous, arranged into two files. Owing to the magnitude of the dataset, only 100% of the total training data is designated for testing when the training and test sets are combined. Seven percent of the dataset is reserved for validation. Various CNN architectures, such as MobileNetv2, efficientNet, and RandomForestClassifier, are employed for analysis. Notably, RandomForestClassifier exhibits the highest accuracy in this scenario. The selected best model undergoes an ablation study based on hyperparameters, with RandomForestClassifier achieving exemplary results: Train Accuracy of 99% and Test Accuracy of 98.37%.