Multi-scale/attention Convolutional Network for Diffuse Large B Lymphoma Segmentation in PET Image

Diffuse large B lymphoma (DLBCL) is one of the most frequent lymphomas, and Positron emission tomography (PET) is an important imaging tool for tumor diagnosis. It can be used to obtain information such as tumor size, location, etc. Clinically, manual segmentation of lesions is time-consuming and labor-intensive. Therefore, in this paper we propose an automatic segmentation model called CBS-APU-Net based on U-Net framework. This model fuses multiple scale and attention mechanisms. During downsampling phase, the CBS module utilizes multi-scale convolution to extract high-dimensional features, and the local receptive field is enlarged using the atrous spatial pyramid pooling at the bottom of the model. The attention mechanism is introduced in the skip connection part so as to enhance the network attention of the lesion area in the feature map. While in the upsampling part, the features after high-dimensional deconvolution are fused with low-dimensional features. Experimental results show that this model achieves the best performance after 500 iterations, and the overall Dice coefficient might reach 0.8189, which outperforms than current models for DLBCL automatic segmentation.