Scholarlink

Ultrasonic imaging plays a vital role in diagnosing thymic lesions, but accurate lesion segmentation remains challenging due to low contrast, blurred boundaries, and high anatomical variability. To address this, we propose KC-UNet++, a semantic segmentation model that integrates a Kolmogorov-Arnold network (KAN) and the CBAM attention mechanism. By combining nested dense skip connections with convolutional attention, the model enables refined feature extraction. The use of spline-based nonlinear modeling and sequence-based feature representation allows KC-UNet++ to capture global context and enhance multi-scale feature learning. Experiments on ultrasound datasets of thymic hyperplasia and tumors demonstrate that KC-UNet++ achieves IoU and Dice scores of 93.0% and 96.3%, outperforming models such as UNet, PAN-UNet, UNet++, LinkNet, PSPNet, and FPN. On average, it improves IoU by 1.1–1.9% and Dice by 0.7–1.3%. This study offers a high-precision, modular approach for the intelligent diagnosis of thymic lesions, with strong potential for broader medical image analysis applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

KC-UNet++: A Thymic Tumor Segmentation Method Based on Kolmogorov-Arnold Network and CBAM

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hao Zheng,
  • Zhonglin Ding,
  • Feng Chen,
  • Wenchao Xia,
  • Hongli Cui,
  • Ruidong Huang,
  • Chuanlei Zhang

摘要

Ultrasonic imaging plays a vital role in diagnosing thymic lesions, but accurate lesion segmentation remains challenging due to low contrast, blurred boundaries, and high anatomical variability. To address this, we propose KC-UNet++, a semantic segmentation model that integrates a Kolmogorov-Arnold network (KAN) and the CBAM attention mechanism. By combining nested dense skip connections with convolutional attention, the model enables refined feature extraction. The use of spline-based nonlinear modeling and sequence-based feature representation allows KC-UNet++ to capture global context and enhance multi-scale feature learning. Experiments on ultrasound datasets of thymic hyperplasia and tumors demonstrate that KC-UNet++ achieves IoU and Dice scores of 93.0% and 96.3%, outperforming models such as UNet, PAN-UNet, UNet++, LinkNet, PSPNet, and FPN. On average, it improves IoU by 1.1–1.9% and Dice by 0.7–1.3%. This study offers a high-precision, modular approach for the intelligent diagnosis of thymic lesions, with strong potential for broader medical image analysis applications.