Official Implementation for "Noise and Anatomy Adaptive Convolution for Efficient Low-Dose CT Denoising"

This project contains redesigned convolution operation implemented using OpenAI Triton.

Key Features

• Modified Convolution operation using Triton: Since it also includes reverse operations, additional learning is possible.

Requirements

• PyTorch (we conducted all experiments using version 2.7.0)
• Triton (automatically installed when you install pytorch)
• LEAP (if using Projector)
• spatial-correlation-sampler
• Donwload trained model files from models

(DnCNN4 & U-Net4 are ours)

Acknowledgements

Our implementation is primarily inspired by Fast and High Quality Image Denoising via Malleable Convolution, and uses LivermorE AI Projector for Computed Tomography (LEAP) framework to perform filtered back-projection (FBP) for reconstruction. We appreciate for their contributions!