Classification of liver lesions in CT images based on LivlesioNet, modified Multi-Scale CNN with bridge Scale method

Liver cancer is one of main origins of death worldwide, and it is the second frequent cause of cancer-related deaths from men and the sixth from women. Deep learning techniques, and in particular convolutional neural networks (CNN) has led to very good performance on a variety of problems, such as image classification, speech recognition, and visual recognition with large-scale annotated datasets. However, collecting medical dataset is difficult and expensive task requires the collaboration of medical experts and researchers. In this paper, we propose a new method for classifying liver lesions based on small dataset. Our proposed method is demonstrated on a small liver lesions dataset of computed tomography (CT) images of 120 patients (30 Hepatocellular carcinoma, 23 metastases, 26 hemangiomas and 41 Healthy). We first use LivlesioNet based on DenseNet to extract feature from the input, in which the model produces the effective feature maps at each stage. Secondly, we propose the modified multi scale convolutional layer (MMS), which extract scale-invariant patterns with a dynamic selection of the convolutional kernels. Based on efforts from LivlesioNet and the modified multi scale convolutional layer, the number of parameters is decreased, and training with small dataset becomes realizable. Then, in order to improve the accuracy, a bridge scale (BS) is proposed to integrate multi-scale spatial features with the aim of removing the redundant features, and adjust weights of the features maps. In addition, after concatenate layer, a fully-connected layer and a SoftMax classifier are connected for further classification. The results indicate that the proposed method provides good results with 97.7% accuracy. It is also indicated that our proposed model achieves the best result in ternary classification.