DCSLK: Combined large kernel shared convolutional model with dynamic channel Sampling.

Area of Science:

• Computer Vision
• Deep Learning
• Medical Image Analysis

Background:

• Convolutional Neural Networks (CNNs) with large kernels face challenges in parameter count and computational complexity.
• Performance gains from increasing kernel size have plateaued, with some methods degrading accuracy.
• Existing methods struggle with spatial feature loss and memory access during channel compression.

Purpose of the Study:

• To develop an efficient deep learning model for computer vision tasks, particularly medical image segmentation.
• To address the parameter and computational complexity issues associated with large convolutional kernels.
• To improve the accuracy of tumor subregion segmentation in medical imaging.

Main Methods:

• Incorporation of a shared parameter mechanism for large convolutional kernels, inspired by human visual processing.
• Synergistic use of large kernels for receptive field expansion and small kernels for fine-grained feature extraction.
• Implementation of a dynamic channel sampling approach to mitigate spatial feature loss during 1x1 convolution channel compression.

Main Results:

• The proposed model demonstrates superior performance across all metrics compared to mainstream Convolutional Neural Networks (CNNs) and Vision Transformers.
• The parameter sharing mechanism effectively reduces model complexity while maintaining the ability to capture spatial relationships.
• Dynamic channel sampling significantly enhances the accuracy of tumor subregion segmentation.

Conclusions:

• The novel approach offers a promising solution for efficient and effective deep learning models in computer vision.
• The methodology provides a new technical strategy for medical image segmentation, outperforming existing architectures.
• This research opens new avenues for developing advanced deep learning techniques in medical imaging.