Identifying subtypes of primary liver tumors based on ConvNeXt and efficient channel-spatial attention mechanism

The preoperative non-invasive differential diagnosis of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) mainly depends on imaging. However, the accuracy of conventional imaging and radiomics methods to distinguish the two is not satisfactory. This paper proposes a convolution neural network model based on a series of pure convolutional neural networks (ConvNeXt), an efficient channel and spatial attention module (ECSAM) and transfer learning, ConvNeXt-ECSAM model, to predict the subtypes of primary liver tumors. The ConvNeXt uses an inverted bottleneck structure to capture the fine-grained features. The ECSAM extracts meaningful attention maps from both of the channel and spatial dimensions. Simulation results on the dataset of preoperative Computed Tomography images showed that for the test dataset, the accuracy and the area under receiver operating characteristic curve of the proposed ConvNeXt-ECSAM model were 0.966 and 0.967, 8.2% and 8.4% on average higher than other four typical deep learning methods, respectively. An ablation study showed that the ECSAM model developed in this paper can improve the prediction performance of ConvNeXt model by an accuracy increase of 1.9%. Our ConvNeXt-ECSAM model can effectively differentiate between HCC and ICC, which has an important clinical significance for preoperative diagnosis of liver tumors.