Medical Image Segmentation Model Based on Multi-scale Dynamic Fusion and Deformable Cross Attention
Yanqing Sun, Lei Wang, Shuqi Liang, Shilong Liu, Shanliang Yang, Lubin Yu, Zhouchao WeiIntroduction:
In recent years, new medical image segmentation methods have been constantly emerging, and their segmentation effects have been significantly improved compared with traditional methods. Medical image segmentation methods based on deep learning have also achieved many excellent advancements. However, they still face problems such as difficulties in multi-scale feature fusion, blurred organ boundaries, and the absence of small targets.
Methods:
This paper proposes an MDFNet model based on multi-scale dynamic fusion and a deformable cross-attention mechanism. This model takes PVTv2 as the backbone network and effectively overcomes the limitations of traditional one-way feature fusion by introducing deformable convolution and bidirectional cross-attention mechanisms, achieving the collaborative expression of local details and global semantics.
results:
The experimental results show that the model has a good segmentation effect on seven medical datasets such as ClinicDB. The Dice coefficient on Kvasir-SEG reached 91.8%, which was 10% higher than that of U-Net, and the number of parameters was only 21.5 million.
Results:
The experimental results show that the model has a good segmentation effect on five medical datasets, such as ClinicDB. The Dice coefficient of Kvasir-SEG reaches 91.8%, which is 10 percentage points higher than that of U-Net, and the number of parameters is only 27 million.
Discussion:
To verify the effectiveness and accuracy, experiments on five popularly used datasets are carefully designed and implemented. Meanwhile, the performance of the proposed method is compared with typical segmentation models, such as the other seven models, like UNet and SANet.
conclusion:
This paper proposes a new medical image segmentation model, MDFNet. This model effectively solves the scale sensitivity problem in medical image segmentation through multi-level dynamic feature interaction and spatio-temporal attention mechanism, providing an effective solution for clinical tasks such as digestive system lesion detection and tumor edge delineation.
Conclusion:
This paper proposes a new medical image segmentation model, MDFNet. This model effectively solves the scale sensitivity problem in medical image segmentation through multi-level dynamic feature interaction and a spatio-temporal attention mechanism, providing an effective solution for clinical tasks such as tumor edge delineation.