An Improved Depth Separable Convolutional Neural Networks for Efficient and Accurate Melanoma Detection

In recent years, Convolutional Neural Networks (CNNs) have achieved significant growth and outstanding success in the field of image classification. This paper explores the performance of a deep separable convolutional neural network optimized based on Google’s proposal, applied to the task of melanoma classification. In addition, several machine learning models, including k-nearest neighbors (K-NN), Gradient Boosting Decision Trees (GBDT), and Adaptive Boosting (AdaBoost), were constructed to evaluate the models from multiple dimensions for industrialisation. In addition to comparing model performance, this paper also innovatively introduces the concept of “model cost-effectiveness”, which provides reference recommendations for model selection in related field tasks considering the future deployment and application scenarios of the models. The experimental results show that the improved deep separable convolutional neural network we constructed performs more comprehensively and favorably on different indicators, while traditional learning models consume less computing and human resources.