Adaptive wavelet enhancement and stratified feature fusion for robust medical image segmentation

Abstract

Accurate medical image segmentation plays an important role in ultrasound triage, colonoscopy screening, and histopathology analysis. However, segmentation performance often degrades substantially when input images are noisy or low-contrast and the model must operate on resource-constrained devices. This challenge mainly arises from two coupled factors. On the one hand, degraded inputs weaken boundary information and interfere with subsequent feature extraction. On the other hand, conventional skip connections with uniform channel allocation cannot simultaneously preserve shallow details and rich deep semantics, thereby limiting the effectiveness of multi-scale feature fusion. To address these issues, we propose UKDW, a lightweight medical image segmentation framework that jointly improves input quality and cross-layer feature interaction under a unified architecture and training protocol. First, we design an Adaptive Wavelet Enhancement Preprocessor (AWEP), which transforms conventional wavelet denoising into an end-to-end trainable, task-driven module. Through a learnable shared soft-threshold, a post-denoising refinement convolution, and fusion weights, AWEP suppresses dataset-level noise while preserving critical structural details. Second, we propose a Dual-Stream Feature Fusion (DSFF) module, which constructs a shallow detail stream and a deep semantic stream with asymmetric channel allocation so that features at different levels receive representational capacities matched to their functional roles. In addition, a semantic-guided attention mechanism emphasizes task-relevant boundaries and suppresses redundant responses, thereby improving cross-scale fusion quality. Experimental results on three heterogeneous public datasets, namely BUSI ultrasound, CVC-ClinicDB colonoscopy, and GLAS histopathology, show that UKDW achieves superior Dice and IoU performance over recent CNN-, Transformer-, and KAN-based methods while maintaining a compact model size of 6.6M parameters in the reported lightweight configuration. These findings suggest that, in lightweight medical image segmentation, jointly addressing input degradation and feature-fusion mismatch is beneficial. It should be noted that the current experiments are based on image-level random splits rather than patient-level, sequence-level, or cross-center validation, and therefore broader clinical effectiveness remains to be further investigated. Code is available at https://github.com/Nafsae/UKDW.