RFPNet: Reorganizing feature pyramid networks for medical image segmentation

Abstract

Medical image segmentation is a crucial step in clinical treatment planning. However, automatic and accurate medical image segmentation remains a challenging task, owing to the difficulty in data acquisition, the heterogeneity and large variation of the lesion tissue. In order to explore image segmentation tasks in different scenarios, we propose a novel network, called Reorganization Feature Pyramid Network (RFPNet), which uses alternately cascaded Thinned Encoder-Decoder Modules (TEDMs) to construct semantic features in various scales at different levels. The proposed RFPNet is composed of base feature construction module, feature pyramid reorganization module and multi-branch feature decoder module. The first module constructs the multi-scale input features. The second module first reorganizes the multi-level features and then recalibrates the responses between integrated feature channels. The third module weights the results obtained from different decoder branches. Extensive experiments conducted on ISIC2018, LUNA2016, RIM-ONE-r1 and CHAOS datasets show that RFPNet achieves Dice scores of 90.47%, 98.31%, 96.88%, 92.05% (Average between classes) and Jaccard scores of 83.95%, 97.05%, 94.04%, 88.78% (Average between classes). In quantitative analysis, RFPNet outperforms some classical methods as well as state-of-the-art methods. Meanwhile, the visual segmentation results demonstrate that RFPNet can excellently segment target areas from clinical datasets.