We provide the `compute_fid_2-5d_ct.py` script that calculates the Frechet Inception Distance (FID) between two 3D medical datasets (e.g., **real** vs. **synthetic** images). It uses a **2.5D** feature-extraction approach across three orthogonal planes (XY, YZ, ZX) and leverages **distributed GPU processing** (via PyTorch’s `torch.distributed` and NCCL) for efficient, large-scale computations.

- **Distributed Processing**

Scales to multiple GPUs and larger datasets by splitting the workload across devices.

- **2.5D Feature Extraction**

Uses a slice-based technique, applying a 2D model across all slices in each dimension.

- **Flexible Preprocessing**

Supports optional center-cropping, padding, and resampling to target shapes or voxel spacings.

Suppose your **real** dataset root is `path/to/real_images`, and you have a `real_filelist.txt` that lists filenames line by line, such as:

You also have a **synthetic** dataset in `path/to/synth_images` with a corresponding `synth_filelist.txt`. You can run the script as follows:

torchrun --nproc_per_node=2 compute_fid_2-5d_ct.py \

--model_name "radimagenet_resnet50" \

--real_dataset_root "path/to/real_images" \

--real_filelist "path/to/real_filelist.txt" \

--real_features_dir "datasetA" \

--synth_dataset_root "path/to/synth_images" \

--synth_filelist "path/to/synth_filelist.txt" \

--synth_features_dir "datasetB" \

--enable_center_slices_ratio 0.4 \

--enable_padding True \

--enable_center_cropping True \

--enable_resampling_spacing "1.0x1.0x1.0" \

--ignore_existing True \

--num_images 100 \

--output_root "./features/features-512x512x512" \

--target_shape "512x512x512"

This command will:

1. Launch a distributed run with 2 GPUs.

2. Load each `.nii.gz` file from your specified `real_filelist` and `synth_filelist`.

3. Apply 2.5D feature extraction across the XY, YZ, and ZX planes.

4. Compute FID to compare **real** vs. **synthetic** feature distributions.

For more details, see the in-code docstring in [`compute_fid_2-5d_ct.py`](./scripts/compute_fid_2-5d_ct.py) or consult our documentation for a deeper dive into function arguments and the underlying implementation.