Hairlines removal and low contrast enhancement of melanoma skin images using convolutional neural network with aggregation of contextual information

Abstract

Skin cancer image preprocessing in dermoscopic images is a fundamental yet challenging task in designing highly efficient computerized diagnostic systems. The presence of hairlines and poor resolution of the skin images poses a significant problem in accurately identifying features for automatic cancer classification. Two approaches are presented here; the first is intensity adjustment-based hair removal (IA-HR) to occlude hairlines by using morphological operators to adjust intensity levels and using nearest neighbour criteria to detect and eliminate hair pixels. A multiscale context aggregation convolutional neural network (MCACNN) is used in the second approach to reduce high-frequency content loss and increase the resolution of dermoscopy images. Evaluation of the proposed methods was conducted using the publicly available dataset HAM10000. For comparison, an additional benchmark dataset H13-Sim was used comprising 13 dermoscopic skin cancer images. Experimental results showed that the proposed methods outperformed state-of-the-art methods using PSNR, SSIM, NIQE, and BRISQUE metrics. It is concluded that the IA-HR method can occlude hairlines effectively while the MCACNN model can improve the resolution of skin cancer images. Consequently, the proposed methods may be used as part of a preprocessing phase in developing a diagnostic system.