Chromosome Image Segmentation Based on KAN Improved U-Net

在产前诊断中，染色体核型分析是检测遗传疾病的金标准，但染色体形态复杂且易交叉粘连，精准分割仍面临挑战。文章提出了新型医学图像分割模型KEU-Net，旨在解决染色体等高复杂度、多尺度目标的精准分割问题。KEU-Net通过融合KAN-PSP模块和跨层级SIMAM注意力机制，显著提升特征建模与边界分割精度。KAN-PSP结合Kolmogorov-Arnold网络的自适应B样条基函数与金字塔池化，实现多尺度特征建模，提升Jaccard指数3.361%；SIMAM注意力机制在编码器深层抑制背景噪声、在解码器浅层增强边缘响应，将边界贴合误差ASD降低19.6%。此外，KAN-PSP与SIMAM通过动态特征协同优化，维持高召回率。实验结果表明，KEU-Net在染色体数据集上的Dice系数为87.074%，较UNet和Attention UNet提升2.57%~3.94%，为医学影像分析提供了高精度、强鲁棒性的解决方案，具有重要的临床应用价值。In prenatal diagnosis, chromosomal karyotype analysis is the gold standard for detecting genetic disorders. However, due to the complex morphology and frequent overlaps of chromosomes, precise segmentation remains a significant challenge. This paper proposes a novel medical image segmentation model, KEU-Net, aimed at addressing the precise segmentation of high-complexity, multi-scale structures such as chromosomes. KEU-Net significantly enhances feature modeling and boundary segmentation accuracy by integrating the KAN-PSP module and cross-hierarchical SIMAM attention mechanism. KAN-PSP combines the adaptive B-spline basis functions of the Kolmogorov-Arnold network with pyramid pooling to achieve multi-scale feature modeling, improving the Jaccard index by 3.361%. The SIMAM attention mechanism suppresses background noise in the deeper encoder layers and enhances edge responses in the shallower decoder layers, reducing the boundary fitting error (ASD) by 19.6%. Furthermore, KAN-PSP and SIMAM work together through a dynamic feature collaborative optimization, maintaining a high recall rate. Experimental results show that KEU-Net achieves a Dice coefficient of 87.074% on the chromosome dataset, outperforming U-Net and Attention U-Net by 2.57%~3.94%. This provides a high-precision and robust solution for medical image analysis with significant clinical application value.