ABSTRACT In the field of small target CT image segmentation, methods based on U‐shaped neural networks have made progress, yet the inherent high noise and low contrast of CT images continue to pose challenges. Traditional U‐Net architectures, relying on CNNs, are limited by their local receptive fields, which hinder their ability to capture long‐range dependencies and contextual relationships essential for identifying subtle defects in noisy conditions. To address this, we propose the Dual‐branch wavelet transform cross‐transformer network (DWTCTNet). Our approach leverages the global dependency modeling of transformers while compensating for their limited sensitivity to high‐frequency details, such as small target edges. By incorporating a dedicated wavelet transform branch, DWTCTNet performs frequency‐domain decomposition to enhance high‐frequency structural information and suppress noise. This branch works collaboratively with the main CNN stream to integrate multi‐scale frequency features with spatial semantics. Additionally, a dense multiplication block (DMB) facilitates deep fusion of encoder features across levels. Experiments on CT image datasets demonstrate that DWTCTNet outperforms existing segmentation methods in accuracy, noise robustness, and handling of complex structures. The code is available from https://github.com/yyx‐ht/DWTCTNet/tree/main .
Jin et al. (Sun,) studied this question.