Preface
This paper proposes Deformable Large Kernel Attention (D-LKA Net), a simplified attention mechanism that uses large convolution kernels to fully understand voxel context, and has demonstrated its effectiveness on academic segmentation datasets (Synapse, NIH Pancreas, and Skin Lesions). Superior performance
Medical image segmentation is significantly improved with the Transformer model, which excels at grasping far-reaching contextual and global contextual information. However, the ever-increasing computational demands of these models (proportional to the squared number of tokens) limit their depth and resolution capabilities. Most current methods process D-volume image data slice by slice (called pseudo-3D), lacking critical inter-slice information, thereby reducing the overall performance of the model. To address these challenges, we introduce the concept of Deformable LargeKernel Attention (D-LKA Attention), a simplified attention mechanism that uses large convolution kernels to fully understand volumetric context. This mechanism is similar to self-attention in runs in the receptive field while avoiding computational overhead. Furthermore, our proposed attention mechanism benefits from deformable convolutions to flexibly distort the sampling grid, enabling the model to appropriately adapt to different data patterns. We design D-LKA Attention 2D and 3D adaptations of , the latter performs well in data understanding across depths. Together, these components shape our novel hierarchical Vision Transformer architecture, D-LKA Net. Our model targets popular medical segmentation datasets (Synapse, Evaluations conducted with leading methods on NIH Pancreas and Skin Lesions) demonstrated its superior performance.
Paper address: https://arxiv.org/abs/2309.00121
Code address: https://github.com/mindflow-institue/deformableLKA
Summarize
The above is what I will talk about today.
Follow the public account CV Algorithm House to learn more about cutting-edge AIGC technology to help you better start a business, publish papers, and find a
job. What should I do if there are no eye-catching projects in the school recruitment?
Join the technical exchange group: Join My WeChat Lh1141755859