FASCANet: Frequency-Aware Spatial Cross-Attention Denoising

A Noise-Supervised Wavelet-Domain Denoising Network for Label-Free Oral Cancer Screening from Autofluorescent Images.

Overview

FASCANet is a noise-supervised framework designed specifically for the low-SNR regime of clinical Autofluorescence Imaging (AFI). It processes images in the wavelet domain to separate low-frequency structural components from high-frequency textures, ensuring that diagnostically critical metabolic signatures are not blurred during noise suppression.

Architecture

The FASCANet architecture is defined by three primary technical pillars:

• Daubechies-2 (db2) Wavelet Decomposition: Uses a single-level 2D discrete wavelet transform to provide a piecewise-linear frequency decomposition. This significantly reduces reconstruction artifacts compared to the simpler Haar basis.
• Spatial Cross-Attention (SCA): A bidirectional coupling mechanism between frequency branches. It generates full-resolution per-pixel attention maps to exchange edge cues and structural context every two residual blocks.
• Wavelet-Domain Residual Learning: The network predicts band-specific corrections (residuals) rather than full image synthesis, preserving the weak endogenous fluorophore signals required for cancer screening.

Benchmark Methods

This repository evaluates FASCANet against the following established denoising baselines:

• FASCANet (Proposed): db2-based wavelet residual network with SCA.
• Noise2Void (N2V): Self-supervised blind-spot masking.
• Neighbor2Neighbor (Ne2Ne): Subsampling-based self-supervision.
• Self2Self: Dropout-based ensemble inference.
• Noise2Same: Blind denoising via self-supervision.
• DIP (Deep Image Prior): Optimization-based reconstruction.
• SwinConv: SCUnet inspired arcehtecture.
• CBM3D: Traditional collaborative filtering.

Project Structure

• main.py: Entry point for benchmarking. Orchestrates dataset splitting, noise injection, and evaluation across random seeds.
• training.py: The training engine. Contains optimized loops for FASCANet and all baselines.
• models.py: Architecture definitions (FASCANet implementation, SCA modules, and UNet).
• utils.py: Utilities for configuration, clinical metrics (PSNR, SSIM, FSIM, VIF, MS-SSIM, FOM), and noise generation.

Installation

pip install torch torchvision numpy opencv-python scikit-image PyWavelets pandas scipy

Configuration

*Update paths and hardware settings in utils.py:

• Config.INPUT_DIR: Point to your AFI dataset.

• Config.OUTPUT_DIR: Define your results directory.

Execution

Run the full benchmarking and ablation suite:

python main.py

Clinical Metrics

The framework evaluates perceptual and structural fidelity using:

• PSNR/SSIM: Standard signal fidelity.

• FSIM: Feature Similarity Index.

• VIF: Visual Information Fidelity.

• MS-SSIM: Multi-Scale Structural Similarity.

• FOM: Pratt’s Figure of Merit for edge preservation.

• Redox Ratio: Verification of $FAD / [NADH + FAD]$ preservation.

Citation

If you utilize this framework, cite the original research: FASCANet: Frequency-Aware Spatial Cross-Attention Denoising for Label-Free Oral Cancer Screening from Autofluorescent Images.