Convolutional Neural Network Classification of Basal Cell Carcinoma in Harmonically Generated Microscopy Images

Basal cell carcinoma (BCC) is the most common form of skin cancer, which could cause local damage of nerves or tissues. Since the tumor growth of BCC is slow and not painful, it could lead to delayed tumor detection and hence necessary subsequent prompt intervention. This paper proposes a computer-aided diagnosis (CAD) method which uses the Gabor filter to extract characteristic scale information according to the characteristic of infected dendritic melanocytes in the third harmonic generation image. Scale information of image which is extracted from Gabor filter allows automatic adjustment of scale range and more accurate segmentation of the infected basal cells in medical images. Subsequently, normal and infected collagen fiber images are used to train convolution neural network (CNN) which are initialized with extracted features as kernels within convolution layers, resulting in high tumor detection accuracy and speed of convergence in harmonically generated microscopy (HGM) images. Experimental results show that this algorithm can accurately classify HGM images, with reduction in time and labor, and thus provides an efficient assisted tool in biomedical image analytics.