Journal Article

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

Li, Yanbin, Li, Zongren, Li, Hongwei, Luo, Shuping

2025

Overview

•Shared Parameter Mechanism of Large Convolutional Kernels: This study proposes a brain tumor segmentation model using shared-parameter large convolutional kernels. It combines 11×11 kernels (capturing global features via broad receptive fields) with 5×5 kernels (extracting fine details). To reduce parameter overload, a sharing mechanism is implemented: central 3×3 regions retain independent parameters for local precision, while peripheral areas share parameters to maintain wide spatial perception. This dual-scale strategy balances computational efficiency with segmentation accuracy, effectively decreasing model complexity while preserving crucial tumor boundary and texture information. The design achieves robust performance through optimized feature extraction across different scales.•Dynamic Channel Sampling Method Enhances Segmentation Accuracy: To enhance segmentation accuracy, this study introduces a dynamic channel sampling method that strategically addresses two critical challenges associated with 1×1 convolutional channel compression: spatial feature information loss and elevated memory access demands. By implementing an adaptive mechanism to dynamically adjust channel sampling strategies during processing, the proposed approach effectively preserves essential spatial features while concurrently optimizing memory utilization. This dual improvement not only mitigates performance degradation caused by rigid compression techniques but also yields a significant enhancement in slice segmentation accuracy, demonstrating the method's capability to balance computational efficiency with feature preservation in medical imaging tasks.•Experimental Validation and Performance Advantages: The model was rigorously validated on BraTs2020, BraTs2024, and Medical Segmentation Decathlon Brain 2018 datasets, outperforming state-of-the-art ConvNet and Transformer architectures in Dice coefficient, Hausdorff distance, and sensitivity. By addressing traditional channel compression limitations, it achieved superior segmentation accuracy and set new benchmarks. The framework’s efficacy in balancing global and fine-grained features enabled precise tumor boundary delineation while maintaining computational efficiency. These results provide critical methodological insights for developing lightweight, high-precision medical image segmentation models. The advancements offer practical solutions to clinical neuroimaging challenges, enhancing diagnostic reliability and paving the way for scalable deployment in resource-constrained healthcare environments. This study centers around the competition between Convolutional Neural Networks (CNNs) with large convolutional kernels and Vision Transformers in the domain of computer vision, delving deeply into the issues pertaining to parameters and computational complexity that stem from the utilization of large convolutional kernels. Even though the size of the convolutional kernels has been extended up to 51×51, the enhancement of performance has hit a plateau, and moreover, striped convolution incurs a performance degradation. Enlightened by the hierarchical visual processing mechanism inherent in humans, this research innovatively incorporates a shared parameter mechanism for large convolutional kernels. It synergizes the expansion of the receptive field enabled by large convolutional kernels with the extraction of fine-grained features facilitated by small convolutional kernels. To address the surging number of parameters, a meticulously designed parameter sharing mechanism is employed, featuring fine-grained processing in the central region of the convolutional kernel and wide-ranging parameter sharing in the periphery. This not only curtails the parameter count and mitigates the model complexity but also sustains the model's capacity to capture extensive spatial relationships. Additionally, in light of the problems of spatial feature information loss and augmented memory access during the 1 × 1 convolutional channel compression phase, this study further puts forward a dynamic channel sampling approach, which markedly elevates the accuracy of tumor subregion segmentation. To authenticate the efficacy of the proposed methodology, a comprehensive evaluation has been conducted on three brain tumor segmentation datasets, namely BraTs2020, BraTs2024, and Medical Segmentation Decathlon Brain 2018. The experimental results evince that the proposed model surpasses the current mainstream ConvNet and Transformer architectures across all performance metrics, proffering novel research perspectives and technical stratagems for the realm of medical image segmentation.

• Publisher: Elsevier Inc,Elsevier Limited,Elsevier
• Subject: Accuracy / Brain cancer / Brain research / Brain tumors / Business metrics / Channel compression / Computer vision / Datasets / Dynamic channel sampling / Humans / Image processing / Image Processing, Computer-Assisted - methods / Information processing / Kernels / Large convolutional kernels / Magnetic Resonance Imaging - methods / Methods / Neural networks / Neural Networks, Computer / Parameter sharing mechanism / Receptive field / Sampling / Tumors