motion_aware_DAJBF

Motion-Aware Deformable Attention-Guided Joint Bilateral Filter (DA-JBF) for 4D CT Denoising

This folder contains the code for our motion-aware extension of the Filter2Noise framework, designed for 4D (3D + time) low-dose CT denoising. The method combines deformable attention-guided joint bilateral filtering with a self-supervised training strategy that exploits temporal motion information.

Method

The pipeline consists of three main components:

1. Demons-based Motion Field Estimation: Multi-scale diffeomorphic registration computes dense displacement fields between consecutive temporal frames, providing motion-compensated neighboring frames and motion magnitude maps as auxiliary inputs.

2. Deformable Attention-Guided Joint Bilateral Filter (DA-JBF): Extends the original Filter2Noise bilateral filter with:

   • A deformable attention-based sigma predictor (DLKASigmaPredictor) that uses multi-scale deformable attention blocks to predict spatially varying filter parameters
   • Learnable sampling offsets that adapt the filter kernel shape to local image content and motion patterns
   • Joint bilateral filtering that incorporates both spatial and learned feature-space range kernels

3. Regularized Self-Supervised Loss: A self-supervised training objective that uses motion-compensated temporal averages as pseudo-targets, with a regularization term that progressively enforces consistency between denoised outputs.

Input Format

The code expects 4D volumes stored as HDF5 files with a volume dataset of shape (T, Z, X, Y) (internally transposed to (X, Y, Z, T)).

Usage

python filter2noise_4d_dajbf.py \
    --clean path/to/clean_volume.h5 \
    --noisy path/to/noisy_volume.h5 \
    --output denoised_4d.h5 \
    --output-dir output_4d

Argument	Default	Description
--clean	(required)	Path to clean 4D volume (HDF5, for evaluation)
--noisy	(required)	Path to noisy 4D volume (HDF5)
--output	denoised_4d.h5	Output denoised volume path
--output-dir	output_4d	Output directory for figures and metrics
--epochs	100	Number of training epochs
--batch-size	16	Training batch size
--lr	0.001	Learning rate
--gamma-max	2.0	Maximum regularization weight for the self-supervised loss
--train-fraction	0.10	Fraction of (z, t) pairs used for training
--seed	77	Random seed

Requirements

In addition to the base Filter2Noise dependencies (requirements.txt in the parent directory), this code requires:

• h5py
• loguru