EFFICIENT 3D TRANSFORMER WITH CLUSTER-BASED DOMAIN-ADVERSARIAL LEARNING FOR 3D MEDICAL IMAGE SEGMENTATION

Real-world application of volumetric medical image segmentation is still challenging due to the domain shift problem and the disability to process volumetric information efficiently by existing algorithms. To address these problems, we propose a 3D Swin Transformer with a pyramidal downsampling strategy to process volumetric information efficiently, dubbed as PDSwin. Specifically, the improved 3D Swin Transformer includes a spatial downsampling strategy that downsamples 2D slices pyramidally according to the spatial relationship, reducing the computation complexity while providing a wider downsampled receptive field. Furthermore, we propose a cluster-based domain-adversarial learning algorithm to attenuate the domain shift problem. The algorithm generates fine-grained cluster-based domains instead of employing center-based domains, ameliorating the domain-adversarial learning performance. We evaluated our model against other competitive models on brain stroke lesion segmentation and prostate segmentation tasks. Extensive experimental results indicated that our proposed model outperforms other models, demonstrating the efficacy of our proposed method.