MDDC: melanoma detection using discrete wavelet transform and convolutional neural network

Abstract

Detection of melanoma in the early stage can provide the patients with more promising treatments. Numerous methods have been proposed for detecting melanoma using computational approaches; however, the application of deep learning in this field is incipient and needs to be assessed further. In this paper, we introduce MDDC, a four-step deep learning-based pipeline for melanoma detection using digital images. MDDC removes image noise using discrete wavelet transform, pre-processes and extends the generalizability using a Gaussian blur filter and image transformation techniques, and finally computes the probability of melanoma in digital images using a convolutional neural network. The model performance is evaluated by considering various CNN architectures and three validation scenarios to find the best CNN architecture. Moreover, we used the transfer learning technique to improve the model performance and generalize it for working on another dataset. MDDC can handle noisy images and is insensitive to image transformations. The validation of MDDC led to 0.96 Accuracy, 0.96 Recall, 0.954 Specificity, 0.99 AUC, and 0.99 AUPR. In addition, further evaluation of the proposed method with state-of-the-art methods verifies that MDDC outperforms other methods. Consequently, MDDC can efficiently and quickly detect melanoma in digital images.