Lesion segmentation in medical images is essential for diagnosis and treatment planning. Two major challenges remain: complex background interference and diverse lesion morphology. Existing encoder‑decoder networks mainly improve feature extraction or decoding strategies, but they lack prior guidance before encoding and feature reconfiguration during encoding, limiting their performance on these issues. To overcome this, we propose FreNet, a visual‑prior‑driven feature reconfiguration framework. FreNet performs pixel‑level reconfiguration before encoding and hierarchical feature reconfiguration during encoding for precise lesion segmentation. To suppress background responses, we introduce an Implicit Prior Neural Network (IPNN) that models a continuous spatial field and leverages visual priors from SAM to preprocess the input image. To handle varied lesion shapes, we design a Dual‑domain Feature Reconfiguration (DFR) module that progressively reconfigures backbone features during encoding. DFR comprises a Frequency Decoupling Module (FDM) that separates features in the frequency domain to improve foreground‑background discriminability, followed by a Spatial Localization Module (SLM) that relocates features spatially after decoupling, enhancing spatial stability. Extensive experiments on nine benchmarks covering three imaging modalities show that FreNet consistently outperforms state‑of‑the‑art methods. On the challenging ETIS dataset, Dice scores improve by 5.0% over the previous SOTA and 7.2% over SAM.
Review: FreNet’s prior‑driven pixel preprocessing and dual‑domain feature reconfiguration effectively mitigate background noise and adapt to morphological variability, demonstrating strong potential across multi‑modal medical segmentation tasks.