DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation (2202.00972v2)

Abstract: Deep learning architecture with convolutional neural network (CNN) achieves outstanding success in the field of computer vision. Where U-Net, an encoder-decoder architecture structured by CNN, makes a great breakthrough in biomedical image segmentation and has been applied in a wide range of practical scenarios. However, the equal design of every downsampling layer in the encoder part and simply stacked convolutions do not allow U-Net to extract sufficient information of features from different depths. The increasing complexity of medical images brings new challenges to the existing methods. In this paper, we propose a deeper and more compact split-attention u-shape network (DCSAU-Net), which efficiently utilises low-level and high-level semantic information based on two novel frameworks: primary feature conservation and compact split-attention block. We evaluate the proposed model on CVC-ClinicDB, 2018 Data Science Bowl, ISIC-2018 and SegPC-2021 datasets. As a result, DCSAU-Net displays better performance than other state-of-the-art (SOTA) methods in terms of the mean Intersection over Union (mIoU) and F1-socre. More significantly, the proposed model demonstrates excellent segmentation performance on challenging images. The code for our work and more technical details can be found at https://github.com/xq141839/DCSAU-Net.

• The paper presents DCSAU-Net, which incorporates a Primary Feature Conservation strategy and Compact Split-Attention block to enhance deep medical image segmentation.
• It employs depthwise separable convolutions and multi-path attention to capture multi-scale features while reducing computational overhead.
• DCSAU-Net outperforms state-of-the-art models with an mIoU of 0.861 and F1-score of 0.916 on challenging datasets, demonstrating strong clinical potential.

DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation

The paper "DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation" introduces an innovative deep learning architecture aimed at improving the efficacy of medical image segmentation tasks. This research addresses current limitations in U-Net architectures, particularly in capturing sufficient feature information at various depths due to the uniform design of downsampling layers in encoders and simplistic convolution stacking.

Methodology

The authors propose DCSAU-Net, a deeper and more compact version of U-Net, incorporating two significant innovations: the Primary Feature Conservation (PFC) strategy and the Compact Split-Attention (CSA) block. The PFC strategy is designed to reduce computational complexity and preserve primary image features with enhanced efficiency. It leverages depthwise separable convolutions with large kernel sizes to improve the receptive field and feature preservation while maintaining fewer parameters and computational overhead.

The CSA block plays a pivotal role in strengthening feature representation across various receptive fields using a multi-path attention mechanism. This block aggregates multi-scale features through distinct convolution paths, thus ensuring robust feature extraction for varying lesion sizes common in medical imaging.

Experimental Evaluation

The proposed DCSAU-Net was evaluated on several challenging datasets, such as CVC-ClinicDB, 2018 Data Science Bowl, ISIC-2018, and SegPC-2021. The model demonstrated superior performance metrics compared to several state-of-the-art (SOTA) models, including mIoU and F1-score, which are crucial for assessing segmentation effectiveness. For instance, on the CVC-ClinicDB dataset, DCSAU-Net achieved an mIoU of 0.861 and an F1-score of 0.916, outperforming traditional U-Net and more recent models like DoubleU-Net and transformer-based architectures.

Implications and Future Directions

The introduction of DCSAU-Net holds substantial implications for medical image segmentation, especially in clinical scenarios requiring precise delineation of pathological structures. Its capability to maintain high segmentation accuracy while being computationally efficient is critical for real-world clinical applications where resources are often limited.

From a theoretical perspective, the model showcases the potential of incorporating attention mechanisms and depthwise separable convolutions in medical image analysis. In future work, enhancing these components or integrating additional neural architecture optimizations may yield even more efficient and accurate segmentation models. Further investigation into diverse clinical imaging modalities and conditions could expand the applicability and robustness of DCSAU-Net.

DCSAU-Net represents a meaningful stride forward in the domain of medical image segmentation, addressing current methodological challenges while providing a strong foundation for future research and application in AI-driven medical diagnostics.