Feature Pyramid Neural Network Based on Receptive Field for Medical Image Segmentation

The automated segmentation of organs in abdominal CT images facilitates physicians in comprehending the spatial distribution of organs and aids in the diagnosis and treatment of diseases. The kidney and spleen occupy an important position in the human body. In this paper, we designed a new feature pyramid neural network architecture based on receptive field to segment kidney and spleen in abdominal CT images. It is a combination of feature pyramid and Res-UNet. We use a set of convolutions to design a feature pyramid and connect it with Res-UN et to enrich its feature extraction. Different from previous studies, we combined different receptive fields in the feature pyramid to extract multi-scale features. We assume that multi-scale features can improve segmentation results. To verify this, we tested this architecture on a public dataset. The experimental results indicate that our method achieved the best results, with average dice of 85.91% and 91.93%, respectively.