Medical images are often characterized by low contrast and blurred object boundaries due to physical imaging device limitations. Significant challenges are thus posed for automatic segmentation. The feature representation of U-KAN is enhanced by the incorporation of the Kolmogorov-Arnold Network (KAN). However, the global context modeling capacity is limited by its basic convolutional units. In addition, the quality of feature fusion and the accuracy of boundary reconstruction are restricted by skip connections with simple concatenation or addition. To address these issues, a model based on the collaboration of feature enhancement and boundary awareness was proposed. A Contextual Feature Enhancement Block was designed as the fundamental encoder-decoder unit. Critical feature responses were strengthened through built-in contextual aggregation attention. A Content-Guided Attention module was introduced to replace traditional skip connections. Adaptive boundary-aware feature fusion was thereby achieved. Experiments were conducted on the BUSI breast ultrasound dataset and the CVC-ClinicDB colonoscopy polyp dataset. In terms of IoU, Dice, and HD95, comparative methods are outperformed by the proposed model. On the BUSI dataset, an IoU of 68.17%, a Dice of 80.79%, and an HD95 of 3.84 are achieved. Compared with the baseline U-KAN, improvements of 4.17% in IoU and 3.20% in Dice are obtained. A reduction of 2.21 in HD95 is also realized. The effectiveness of the collaborative mechanism is thus validated.