Fig 1.
(a) The overall architecture of our proposed SA-UMamba. (b) RSSS block is the main construction block of SA-UMamba, integrating the VSSM and RFAC modules for enhanced feature extraction. (c) VSSM is the crucial module in the Mamba model for extracting visual features, with SS2D as its core operation.
Fig 2.
(a) SS2D expansion operation; (b) Core component of Mamba (S6); (c) SS2D merging operation.
Fig 3.
RFAC module, which dynamically determines the importance of each feature in the receptive field.
Table 1.
Segmentation results for various methods conducted on Synapse dataset. (The best results are highlighted in bold, and the second best are underlined.)
Fig 4.
Visualization of segmentation results for different methods on the Synapse dataset.
Table 2.
Segmentation results for various methods conducted on ISIC17 and ISIC18 datasets. (The best results are highlighted in bold, and the second best are underlined.)
Table 3.
Segmentation results for CVC-ClinicDB dataset. (The best results are highlighted in bold, and the second best are underlined.)
Table 4.
Synapse dataset segmentation results for differently scaled pre-trained weights with the same backbone. (The best results are highlighted in bold.)
Table 5.
Synapse dataset segmentation results for different RSSS block design choices. (The best results are highlighted in bold.)
Table 6.
Segmentation results for different attention calculation methods on the Synapse dataset. (The best results are highlighted in bold.)
Table 7.
Ablation study on the number of RSSS blocks in encoder and decoder. (The best results are highlighted in bold.)
Table 8.
Comparison of model Params(M), FLOPs(G). These test results were obtained on a single RTX3090. (The best results are highlighted in bold.)