Improved Mask R-CNN with Deformable Convolutions for Accurate Liver Lesion Detection in Multiphase CT Images

Abstract

Multiphase computed tomography (CT) is widely used for the diagnosis of liver lesions. Accurate detection of lesions in multiphase CT images is important in the diagnosis of liver cancer. However, to detect liver lesion from multiphase CT images, the multiphase images must be registered accurately as a preprocessing step, which is complicated by liver deformation. Therefore, we herein propose an improved mask R-CNN (Region with Convolutional Neural Network), in which deformable convolution modules are used for feature extraction, unlike the use of conventional convolution modules in mask R-CNN. Experimental results show that the proposed method performs better compared to the baseline (i.e., the original Mask R-CNN with ResNet-50). In particular, our method yields an overall AP50 of 4.8. In terms of AP (Average Precision) and AP75, our method yields values that are 1.6 and 2.6 higher than those of the original mask R-CNN, respectively. Then in the mask AP, our method yields values that are 1.8, 3.2, and 1.1 higher than those of the original mask R-CNN, respectively.