MRFMA: A hybrid paradigm integrating multi-receptive field network with mediator attention for 3D multi-organ segmentation

Multi-organ segmentation has become a critical task in medical image analysis, and a precise understanding of anatomical structures is crucial for advancing disease diagnosis, treatment planning and prognosis. Existing three-dimensional (3D) multi-organ segmentation algorithms usually combine 3D Convolutional Neural Networks (CNNs) with Transformers, for the purpose of capturing local and global features. However, traditional CNNs with a fixed size of receptive field struggle to adapt to the diverse scales and long-range spatial relationships of multiple organs. Transformers have been widely used to establish dependencies on global information, despite this, they greatly increase the computational complexity. To mitigate these challenges, we propose a hybrid paradigm, called Multi-Receptive Field Network with Mediator Attention (MRFMA), to boost the representation quality for robust multi-organ segmentation across diverse imaging modalities. In MRFMA, the novel multi-receptive field depthwise convolutional module adeptly preserves the inherent inductive biases of convolution while enhancing the network’s capacity to model both fine-grained local patterns and long-range contextual relationships across anatomically disparate organs. Besides, a novel attention mechanism called Mediator Attention is developed to establish dependencies on global information. Mediator attention avoids the direct similarity calculation of query (Q) and key (K) by introducing the Mediator tokens, and thus dramatically decreases the computational cost. The proposed MRFMA is tested on the MM-WHS 2017 CT and MR dataset, FLARE 2021 dataset as well as the BTCV dataset, achieving the average Dice scores of 93.7 %, 82.2 %, 93.0 % and 83.2 % respectively. Extensive experimental results prove that our proposed method achieves superior performances in comparison with state-of-the-art (SOTA) methods. Our code will be released viahttps://github.com/jiatong0925/MRFTA.