LInFBP: Continuous Filtered Backprojection by Learnable Interpolation Network

This repository contains the official implementation of LInFBP, a lightweight and plug-and-play continuous representation framework designed to mitigate discretization and interpolation errors in CT backprojection.

---

Project Structure

The project is organized as follows:

.
├── Datasets/                # Data management
│   ├── cal_fsim.py          # FSIM calculation
│   └── datasets.py          # Data loading and augmentation logic
│   └── imageProcess.py      # Data pre-processing and post-processing
│   └── utils.py             # Helper functions
├── Model/                   # Core architecture implementations
│   ├── backProjNet.py       # Base backprojection network with nearest interpolation
│   ├── backProjNet_L.py     # Proposed L-LInFBP (Linear-basis variant)
│   ├── backProjNet_F.py     # Proposed F-LInFBP (Fourier-basis variant)
│   ├── backProjNet_linear.py# Standard linear interpolation backprojection
│   ├── backProjNet_cubic.py # Standard cubic interpolation backprojection
│   ├── DICDNet.py           # Original DICDNet implementation
│   ├── DICDNet_L.py         # DICDNet enhanced with L-LInFBP
│   ├── DICDNet_F.py         # DICDNet enhanced with F-LInFBP
│   ├── iRadonMap_Net.py     # Original iRadonMap implementation
│   ├── iRadonMap_Net_L.py   # iRadonMap enhanced with L-LInFBP
│   ├── iRadonMap_Net_F.py   # iRadonMap enhanced with F-LInFBP
│   ├── interpolate.py       # Continuous representation & interpolation logic
│   └── model_fbp_*.py       # End-to-end FBP wrappers for various interpolations
├── Solver/                  # Execution scripts
│   ├── train.py             # Training pipeline
│   ├── test.py              # Evaluation pipeline
│   └── pixelIndexCal.py     # CUDA-based coordinate calculation
├── Utils/                   # Helper functions
│   ├── initFunction.py      # Initialization functions
│   └── initParameter.py     # Hyperparameters
├── eval_memory.py           # Efficiency & Memory benchmark script
└── main.py                  # Entry point

Download Precomputed Indices

For 100-view CT reconstruction, LInFBP requires precomputed coordinate indices to accelerate the backprojection process. You can either download the precomputed file or generate it locally.

Please download the 100-view index file from the link below and place it in the Results/ directory:

• Download Link: [Link]
• File Name: indices_100view.dat

Alternatively, you can generate the indices directly by running the provided script. This will automatically compute and save indices_100view.dat into the Results/ folder:

python generate_indices.py

Usage

The execution logic of this project is centralized in Solver/initParameter.py and main.py. Please follow the steps below to configure your environment.

1. Training a Model

To start a new training session:

1. Configure Parameters: Open Solver/initParameter.py and set:
   • net_name: Set to your target model (e.g., LInFBP_L, DICDNet_L).
   • isTrain: Set to True.
   • reload_model: Set to False.
2. Verify Entry Point: Open main.py and ensure the model import corresponds to your net_name selection.
3. Run:

   python main.py

2. Evaluation / Testing

To evaluate a pre-trained model and generate PSNR, NMSE, and FSIM metrics:

1. Configure Parameters: Open Solver/initParameter.py and set:
   • net_name: Must match the architecture of the checkpoint.
   • isTrain: Set to False.
   • reload_model: Set to True.
2. Run:

   python main.py

3. Efficiency Benchmarking

To measure Model Size, FLOPs, Running Time, and Peak Memory usage:

python eval_memory.py