Weak-prior dual cognitive attention lightweight network for abdominal multi-organ segmentation

Automated segmentation of organ regions (ASOR) in abdominal images enjoys high time and labor efficiency, yet it is still very challenging due to the patient-specific variations and complex characteristics of abdominal organs. Existing based ASOR methods suffer from the complex network structure and extensive anatomical information processing. Motivated by this discovery, this paper proposes a novel and effective weak-prior dual attention lightweight segmentation network with the following three-fold ideas: (a) the asymmetric UNet structure is adopted for effectively reducing the network size via the utilization of different residual convolution blocks in the encoder and decoder, (b) the lightweight dual cognitive attention module enhances pixel-level feature representation without excessive memory consumption or matrix operations, and (c) the weak-prior aware module overcomes the difficulties of incorporating strong anatomical prior information (position or shape) into the model by using weak prior (relevance and volume ratio) to guide the attention collaboratively. Evaluations of our method are conducted on open AbdomenCT-1K, AMOS2022 and Synapse datasets. Extensively empirical results demonstrate significant efficiency gains as well as highly competitive segmentation accuracy (the codes are public at https://github.com/feijianli/WDCAL-Net).