基于VMamba-CNN混合的结直肠癌切片图像分割

Colorectal Cancer Slice Image Segmentation Based on VMamba-CNN Hybrid

摘要: 该研究提出一种基于VMamba和卷积神经网络(CNN)混合架构的结直肠癌(CRC)病理切片图像分割方法VMDC-Unet，旨在解决传统方法在处理肿瘤异质性、复杂背景及模糊边界时的不足。该方法通过融合VMamba模型的长距离依赖处理能力和CNN的局部特征提取优势，引入改进的ConvNext模块以增强细粒度特征提取，并设计局部自注意力机制优化跳跃连接的特征融合效率。实验结果表明，在SJTU_GSFPH和Glas数据集上，VMDC-Unet的分割精度与泛化能力均优于其他基准模型，消融实验进一步验证了各模块的有效性。该工作为医学图像分割提供了多模型协同的新思路，其结合全局依赖建模与局部特征强化的策略，为CRC精准诊疗提供了可靠的技术支持。

Abstract: This study proposes a VMDC-Unet method based on a hybrid architecture of VMamba and convolutional neural network (CNN) for colorectal cancer (CRC) pathological image segmentation, aiming to address the limitations of traditional methods in handling tumor heterogeneity, complex backgrounds, and blurred boundaries. The method integrates VMamba’s long-range dependency modeling capability with CNN’s local feature extraction strength. It introduces an enhanced ConvNext module to improve fine-grained feature representation and designs a local self-attention mechanism to optimize feature fusion efficiency in skip connections. Experimental results demonstrate that VMDC-Unet outperformed baseline models in segmentation accuracy and generalization capability on both SJTU_GSFPH and Glas datasets. Ablation studies further verified the effectiveness of each component. The work provides a novel multi-model collaboration strategy for medical image segmentation, where the combination of global dependency modeling and local feature enhancement offers reliable technical support for precise CRC diagnosis and treatment.