MSDSC: A Multistage Deep Learning-based Skin Cancer Classifier

Abstract

One of the most frequent types of cancer in the globe is skin cancer which is a complicated public health issue. A sample of tissue from the skin lesion confirms the diagnosis of skin cancer. However, before making a definitive diagnosis, professionals detect specific signs that can be considered as early diagnosis. An early skin cancer diagnosis is prone to mistakes due to specialists’ lack of knowledge and similarities with other illnesses. This study presents a multistage deep learning-based skin cancer classifier (MSDSC) to give early detection of melanoma and non-melanoma skin lesions. The first stage is the pre-processing of samples to remove the hair surrounding the skin lesion. Since the dataset is unbalanced, the second stage was creating synthetic images using the Generative Adversarial Networks (GAN) model. This stage is followed by an attention-based U-Net model that gener- ates masks for regions of interest to eliminate the background. Lastly, two different types of EfficientNet are trained to classify skin lesions using segmented images. The proposed framework was tested using the International Skin Imaging Collaboration dataset (ISIC) and the results showed that it outperformed the state-of-the-art studies achieving an accuracy of 0.96, F1-score of 0.91, recall of 0.95, and precision of 0.88.

Melanoma image synthesis: a review using generative adversarial networks

“…A novel GAN model with four transposed convolutional layers and a leaky rectified linear unit (LReLU) as an activation function in generator was proposed [45] to avoid vanishing gradients and dying neurons problems in ReLU. The generator network also includes a convolutional layer with the sigmoid activation function and the discriminant consists of five convolutional layers with LReLU as an activation function, followed by a dense layer for classification with sigmoid function.…”