DenseNet model with RAdam optimization algorithm for cancer image classification

Abstract

Application of deep learning algorithms to medical images recognition can improve diagnostic accuracy and efficiency. In recent years, computer-aided diagnosis has attracted the attention of a large number of researchers. The introduction of image processing in medicine is a important method to reduce unnecessary manual diagnosis costs and promote disease classification and detection. In this paper, we propose a novel method for metastatic cancer image classification which uses Densely Connected Convolutional Networks, Rectified Adam optimization algorithm, and focal loss. DenseNet can effectively capture the important features hidden in images. And RAdam optimization algorithm Radam is robust for model training. Our dataset is provided by the Kaggle competition, which is the modified version of the PatchCamelyon (PCam) benchmark dataset. The dataset packs the clinically relevant problem of metastasis detection into a straight-forward binary image classification problem. The experiments shows our approach can effectively identify metastatic cancer in small image patches which are taken from larger digital pathology scans on the dataset. And experimental results indicate that our proposed model is significantly better than Resnet34, Resnet50, Vgg19. The effectiveness of the DenseNet Block, Rectified Adam, focal loss is also verified.