Basal Cell Carcinoma ; Deep Learning ; Convolution Neural Network ; Skin Cancer ; Segmentation
Abstract
A steady increase in incidence of basal cell carcinoma and other skin cancers has been reported in a number of epidemiologic studies conducted in Korea. In this study, we demonstrated a pipelined deep neural network-based model classifying the Korean basal cell carcinoma vs. normal nevus, using photographic images taken in various lighting, angles, and zoom to develop an early-detection diagnosis kits for malignant skin lesions. The original data set consisted of 1,200 photographic images-600 of basal cell carcinoma and 600 of normal nevus. Our study focused on the effectiveness of pipelining two different DNN-based models for this classification task. The performance of the pipeline was compared to the classification-only method based on four metrics: sensitivity, specificity, AUC, and accuracy. Various sizes of the training set were evaluated (n = 50, 100, 300, 600, 840) to assess the performance of each method. The classification-only model reached an AUC of .93. Closer inspection on the incorrectly-classified test points revealed that the model tended to mislabel noisy images. With the segmentation-classification model, the segmentation step allowed tighter cropping around the lesions, cutting out the distractions. Consequently, we found that for a given small dataset, the segmentation-classification model performed better than the classification-only model. This is a pioneering research in examining a deep learning system for the Korean-specific photographic images. While our study focused primarily on the Korean basal cell carcinoma, our data demonstrated the potential in cancer detection in the clinic where a sufficient number of quality training data are not available. Although we utilized noisy images in low resolution as input data, our findings demonstrated that the pipelining procedures can provide a promising system to deal with data having such limitations.