An Efficient Dual-Pooling Channel Attention Module for Weakly-Supervised Liver Lesion Classification on CT Images

Abstract

The classification of liver lesions on computed tomography (CT) images is vital for early liver cancer diagnosis. To aid physicians, several methods have been proposed to perform that task automatically. These methods mostly use regions of interest (ROIs) that are manually or automatically selected from liver CT slices as training and testing data. However, the manual selection of ROIs requires a lot of time and effort from experienced physicians making the number of available fully labeled datasets limited. When it comes to automatic ROIs selection using segmentation methods, some incorrect mask predictions that might occur can affect the quality of the selected ROIs causing the model for the final task to be trained improperly. In this paper, we explore the potential of using the whole liver CT slice in combination with the lesion type label for weakly-supervised lesion classification. We exploit the ability of attention mechanisms to lead convolutional neural networks (CNNs) to focus on important regions in the input image. Then, we propose a lightweight attention module modified from the Efficient Channel Attention (Eca) module, called Efficient Dual-Pooling Channel Attention (EDPca). Our experimental results show that attention mechanisms can significantly boost the performance of CNNs for the task of weakly-supervised liver lesion classification, bringing the results closer to that of the supervised task. Besides, our proposed module achieves better classification results than the other channel attention ones. Moreover, the combination of it and the spatial attention module from CBAM module outperforms other methods.