The proposed metric is based on frequency features statistics.
The provided source-code runs in python. The required packages can be installed using conda using the following:
conda create -n public_metric python=3.8
conda activate public_metric
conda install -c anaconda numpy scipy imageio matplotlib scikit-learn pandas
Optionally, for generating KDE cosine plots, seaborn can be installed:
conda install -c anaconda seaborn
The code has been tested with the packages described in requirements.txt.
The spatial metric can be computed by running the following command:
python spatial_metric.py --num_images 1000 --path_to_reals <path-to-folder-containing-real-pngs> --path_to_fakes <path-to-folder-containing-fake-pngs> --results_dir ./public_metric/
The frequency metric can be computed by running the following command:
python freq_metric.py --num_images 1000 --path_to_reals <path-to-folder-containing-real-pngs> --path_to_fakes <path-to-folder-containing-fake-pngs> --results_dir ./public_metric/
The terms of use for this software are governed by the contents of the accompanying LICENSE file.
Zhikai Yang, et al. "Efficient Generation of Synthetic Breast CT Slices By Combining Generative and Super-Resolution Models" MICCAI DeepBreast Workshop (2024) Zhikai Yang, et al. "Combining Diffusion and Super-resolution Models for Efficient Generation of Synthetic Breast Images" https://epos.myesr.org/poster/esr/ecr2024/C-24121 Deshpande, Rucha, et al. "Report on the AAPM Grand Challenge on deep generative modeling for learning medical image statistics." ArXiv (2024).