Precise cervical cancer cell boundary denoising and segmentation with adaptive wavelet-spectral enhancement

Abstract

Accurate segmentation of cell nuclei in cervical cytology images is crucial for automated cervical cancer screening, yet existing methods struggle with blurred boundaries, noise-induced degradation, and topologically implausible predictions. The current research proposes Cell-Seg Tool, a novel triplet-branch diffusion AI tool that synergistically integrates three innovations to address these limitations. The Wavelet-Enhanced Contour Refinement Branch employs a learnable multi-scale discrete wavelet transform with adaptive coefficient attention to dynamically enhance boundary features across horizontal, vertical, and diagonal orientations. The Adaptive Spectral Noise Suppression module performs dual-domain processing using DCT-based filtering and uncertainty-guided fusion, coupled with bidirectional anchor semantic feedback to couple cross-branch information. The Topology-Aware Hybrid Loss integrates a focal Tversky loss, a persistent homology loss, a directional boundary loss, a skeleton completeness loss, and a diffusion-noise MSE loss for multi-objective optimization. Comprehensive experiments on multiple datasets demonstrate superior performance, achieving 94.45% Dice coefficient and 19.2% reduction in boundary localization error compared to state-of-the-art methods. Unlike prior work that applies these techniques independently, this work demonstrates that their adaptive, synergistic integration within a diffusion-based framework yields substantial improvements in boundary accuracy and topological correctness.