U-Next: A Novel Convolution Neural Network With an Aggregation U-Net Architecture for Gallstone Segmentation in CT Images

Abstract

The incidence of cholelithiasis is more than 10% in the natural population. It is a common and frequently occurring disease worldwide. The lesions are predominantly sand-like and are difficult to distinguish in medical images. Although ultrasound is often the first-line technique for diagnosing cholelithiasis, CT plays an important role in diagnosing complications related to gallstones. To the best of our knowledge, no effective method has been proposed for segmenting CT images of gallstones. To address this difficulty, we have proposed a novel deep learning segmentation model, U-NeXt, which can automatically provide the importance of different shapes and sizes of target structures in images. a) We proposed attention up-sampling blocks for up-sampling. b) We also proposed a spatial pyramid pooling of skip connections, which is called SkipSPP. A series of convolution layers are connected by jumping connections to generate multi-scale features. c) We employed a dense connection to many details of the model to extract more features. In addition, we contributed a new data set for gallstone segmentation, which is co-labelled by three chief physicians from the hepatobiliary department and the medical imaging department of Shandong Province Third Hospital, China. The data set included 5,350 images from 726 patients. Compared with the previous best CT image segmentation method for gallstones, the intersection over union (IoU) improved by 24%, and compared with the medical image segmentation baseline model, the IoU improved by 7%, with a satisfactory segmentation effect. We extended U-NeXt to the nuclei segmentation task, showing that the model has good generalizability.

Keywords

• Image segmentation
• artificial intelligence
• neural networks
• medical diagnosis