Adaptive frequency-domain enhanced deep model driven by heterogeneous networks for medical image segmentation

Abstract

Accurate medical image segmentation necessitates precise localization of global structures and local boundaries due to the high variability in lesion shapes and sizes. However, existing models are limited by conventional spatiotemporal features and single-network architectures, which restrict the simultaneous captures of semantic information and boundary details, thereby challenging generalizable medical image segmentation. To overcome these limitations, we propose a heterogeneous network-driven adaptive frequency-domain enhanced deep model(AFDSeg). First, we introduce the Frequency Domain Adaptive High-Frequency Feature Selection(FAHS) module, which adaptively extracts high-frequency features to enhance contour and detail representation while integrating spatiotemporal and frequency-domain features for improved consistency. Additionally, Prototype-Guided Low-Frequency Feature Aware(PFLA) and Local High-Frequency Salient-Feature Denoising (LHSD) modules are developed, which extract discriminative low-frequency features while suppressing local noise in high-frequency components, thereby facilitating efficient multi-scale feature fusion. Furthermore, the Multi-Level Prototype Feature Refinement(MPFR) Module is introduced to align low- and high-dimensional features during decoding and enhance semantic consistency. Finally, a heterogeneous network framework capable of accommodating multiple network architecture for medical image segmentation is proposed. Our method achieves mDice scores of 93.91%, 88.64%, 90.70%, 91.27%, and 81.38% on the Kvasir-SEG, BUSI, ISIC-2017, ACDC, and Synapse datasets, respectively, and attains 92.09%, 93.50%, and 83.92% in cross-domain experiments on three unseen datasets (Kvasir Capsule-SEG, BUS42, and M&Ms). Our approach consistently outperforms state-of-the-art methods on both benchmark and cross-domain datasets. Extensive quantitative and qualitative experiments demonstrated that AFDSeg accurately segments global structures and local details while maintaining superior generalization, underscoring its clinical significance. The Code is available at https://github.com/promisedong/AFDSeg.