Research on Multi-Objective Optimization of Medical Image Segmentation Based on Frequency Domain Decoupling and Dual Attention Mechanism

Abstract

Medical image segmentation faces the challenge of balancing multiscale anatomical structure modeling and computational efficiency. To address this issue, this paper proposes a “Frequency-Attentive Multi-Hierarchical Network for Medical Image Segmentation” (FreqAtt-MultHier-Net), aiming to achieve synergistic optimization of accuracy, efficiency, and robustness. The core innovations of this paper include: A dual-frequency block (DFB), which decouples high-frequency (detail) and low-frequency (semantic) features through a learnable channel splitting mechanism, and enhances multiscale representations through cross-frequency interaction and dynamic calibration. A multiscale dual-attention fusion block (MSDAFB), which couples channel-spatial dual attention with multi-kernel convolutions to suppress background noise and strengthen local–global contextual fusion. A lightweight ConvMixer module that replaces Transformers with sublinear computational complexity to achieve efficient long-range dependency modeling. In tasks involving cell contour, cell nucleus, lung cancer, skin cancer, liver tumor segmentation and retinal vessel segmentation Task, our model achieves dice similarity coefficients (DSCs) of 95.64%, 92.74%, 83.63%, 85.96%, 85.86% and 84.26%, respectively, while reducing parameter count (25.48 M) and computational cost (5.84 G FLOPs) by 75.9%–84.9% compared to Transformer-based architectures. Ablation experiments validate the independent contributions of each module, with frequency-domain decoupling improving high-frequency detail retention by 18.8% and lightweight design reducing FLOPs by 78.3%. FreqAtt-MultHier-Net provides a high-precision, low-redundancy general solution for medical image segmentation, with potential for low-power clinical deployment. The code is available at the following URL: https://github.com/wu501-CPU/FreqAtt-MultHier-UNet.