KTUNet: A hybrid CNN-Transformer with kolmogorov-arnold convolution for thyroid nodule segmentation in ultrasound images

Objective: Accurate segmentation of thyroid nodules in ultrasound images is essential for facilitating automated diagnosis and personalized treatment planning. Despite the promise of deep learning, existing models face significant challenges when applied to thyroid ultrasound image segmentation. Convolutional Neural Networks (CNNs) often require very deep architectures to model complex nonlinear patterns, while Transformers demand extensive training data, making them less effective in data-restricted scenarios. Methods: To address these challenges, we propose KTUNet, a novel hybrid model that integrates Kolmogorov-Arnold Convolution (KAC) modules with CNNs and Transformers. The KAC modules enable efficient local feature extraction and complex nonlinear modeling, while Transformers capture long-range dependencies. This synergistic design combines the strengths of all three components, resulting in a robust and adaptable framework for thyroid ultrasound image segmentation. Results: Experimental evaluations on the TN3K dataset and a private dataset demonstrate the superiority of KTUNet. The model consistently outperforms state-of-the-art (SOTA) methods, achieving higher segmentation accuracy and boundary precision. For instance, KTUNet achieves a Dice score of 82.49% on the TN3K dataset, surpassing traditional models such as UNet and advanced methods like TransUNet. On the private dataset, KTUNet further consolidates its performance advantage with even higher segmentation metrics. Conclusion: KTUNet represents a significant advancement in thyroid ultrasound image segmentation by addressing the limitations of current deep learning models. Its innovative integration of KAC modules with CNNs and Transformers demonstrates improved accuracy, robustness, and boundary precision, offering a valuable tool for automated thyroid nodule diagnosis and treatment planning.