Generative Modeling for Interpretable Failure Detection in Liver CT Segmentation and Scalable Data Curation of Chest Radiographs - Official Repository
Generative Modeling for Interpretable Failure Detection in Liver CT Segmentation and Scalable Data Curation of Chest Radiographs
M. Woodland, M. Al Taie, C. O'Connor, O. Lebimoyo, J.P. Yung, P.E. Kinahan, C.D. Fuller, B.C. Odisio, A.B. Patel, & K.K. Brock
Abstract
Purpose: To leverage generative modeling-based anomaly detection in two radiological applications, namely, providing interpretability for model failure prediction and enabling scalable data curation for large repositories.
Materials and Methods: This retrospective study utilized 3,306 CT scans for failure prediction, divided patient-wise into training, baseline test, and anomaly (attributes that cause liver segmentation failures—e.g., needles, ascites) test datasets. For data curation, 112,120 ChestX-ray14 radiographs were used for training, and 2,307 radiographs from The Medical Imaging and Data Resource Center for testing, categorized as baseline or anomalous based on attribute alignment with ChestX-ray14. StyleGAN2 networks modeled the training distributions. Test images were reconstructed with backpropagation and scored using mean squared error (MSE) and Wasserstein distance (WD). Scores should be large for anomalous images since StyleGAN2 cannot model unseen attributes. Area under the receiver operating characteristic curve (AUROC) assessed anomaly detection, i.e., differentiating the baseline and anomaly datasets. The proportion of highest-scoring patches containing needles or ascites assessed anomaly localization (interpretability). Permutation tests determined statistical significance. The code and ChestX-ray14 weights are available at .
Results: StyleGAN2 did not reconstruct anomalous attributes (e.g., needles, ascites), enabling the unsupervised detection of attributes that cause model failure or deviate from ChestX-ray14—mean (± standard deviation) AUROCs of 0.82 (±0.16) and 0.82 (±0.11), respectively. 71% (±3%) of the needles and ascites were localized. WD outperformed MSE on CT (p<.001), while MSE outperformed WD on radiography (p<.001).
Conclusion: Generative models detected anomalous image attributes, demonstrating promise for model failure detection interpretability and large repository data curation.
This work has been submitted to Radiology: Artificial Intelligence for possible publication. It is an expansion of the extended abstract titled StyleGAN2-based Out-of-Distribution Detection for Medical Imaging1, accepted to the Medical Imaging Meets NeurIPS workshop at NeurIPS 2022 (abstract, preprint). It was extended with the anomaly localization evaluation, the data curation application, and the generative modeling evaluation.
This code was built with Python 3 with package versions listed in the requirements.txt file.
We also provide a Docker container that can be built by running the following command in this directory.
docker build --tag gan_anom_detect .
Our Python scripts can be run directly after entering the Docker container
docker run -it --rm -v $(pwd):/workspace gan_anom_detect /bin/bash
and providing the appropriate Python commands arguments or by editing the bash scripts with your arguments. The bash script will run the container and provide the arguments to the script for you. You will need to edit the above command or the bash scripts if you want a directory other than the current directory mounted.
To use the StyleGAN2-ADA2 fork submodule, you'll need the following container:
docker build --tag sg2ada:latest stylegan2-ada-pytorch/.
To use the StudioGAN3 fork submodule, you can pull the following container:
docker pull alex4727/experiment:pytorch113_cuda116
To use our bash scripts, you'll need the following Docker container which updated the PyTorch version:
docker build --tag studiogan:latest PyTorch-StudioGAN/.
Lastly, the following Docker container is compatible with the frd-score Python package4.
docker build --tag frd:latest frd/.
Train a StyleGAN2-ADA1 model with the StyleGAN2-ADA repository (forked) by providing the --outdir and --data arguments to the train.sh script.
The provided Optional Arguments in the script were the hyperparameters used to train models in our study.
While the official repository and this study works with PNGs, you can train on NiFTI files using the nifti branch of our fork.
./bash_scripts/train.sh
Usage: train.py [OPTIONS]
Train a GAN using the techniques described in the paper "Training
Generative Adversarial Networks with Limited Data".
Examples:
# Train with custom dataset using 1 GPU.
python train.py --outdir=~/training-runs --data=~/mydataset.zip --gpus=1
# Train class-conditional CIFAR-10 using 2 GPUs.
python train.py --outdir=~/training-runs --data=~/datasets/cifar10.zip \
--gpus=2 --cfg=cifar --cond=1
# Transfer learn MetFaces from FFHQ using 4 GPUs.
python train.py --outdir=~/training-runs --data=~/datasets/metfaces.zip \
--gpus=4 --cfg=paper1024 --mirror=1 --resume=ffhq1024 --snap=10
# Reproduce original StyleGAN2 config F.
python train.py --outdir=~/training-runs --data=~/datasets/ffhq.zip \
--gpus=8 --cfg=stylegan2 --mirror=1 --aug=noaug
Base configs (--cfg):
auto Automatically select reasonable defaults based on resolution
and GPU count. Good starting point for new datasets.
stylegan2 Reproduce results for StyleGAN2 config F at 1024x1024.
paper256 Reproduce results for FFHQ and LSUN Cat at 256x256.
paper512 Reproduce results for BreCaHAD and AFHQ at 512x512.
paper1024 Reproduce results for MetFaces at 1024x1024.
cifar Reproduce results for CIFAR-10 at 32x32.
Transfer learning source networks (--resume):
ffhq256 FFHQ trained at 256x256 resolution.
ffhq512 FFHQ trained at 512x512 resolution.
ffhq1024 FFHQ trained at 1024x1024 resolution.
celebahq256 CelebA-HQ trained at 256x256 resolution.
lsundog256 LSUN Dog trained at 256x256 resolution.
<PATH or URL> Custom network pickle.
Options:
--outdir DIR Where to save the results [required]
--gpus INT Number of GPUs to use [default: 1]
--snap INT Snapshot interval [default: 50 ticks]
--metrics LIST Comma-separated list or "none" [default:
fid50k_full]
--seed INT Random seed [default: 0]
-n, --dry-run Print training options and exit
--data PATH Training data (directory or zip) [required]
--cond BOOL Train conditional model based on dataset
labels [default: false]
--subset INT Train with only N images [default: all]
--mirror BOOL Enable dataset x-flips [default: false]
--cfg [auto|stylegan2|paper256|paper512|paper1024|cifar]
Base config [default: auto]
--gamma FLOAT Override R1 gamma
--kimg INT Override training duration
--batch INT Override batch size
--aug [noaug|ada|fixed] Augmentation mode [default: ada]
--p FLOAT Augmentation probability for --aug=fixed
--target FLOAT ADA target value for --aug=ada
--augpipe [blit|geom|color|filter|noise|cutout|bg|bgc|bgcf|bgcfn|bgcfnc]
Augmentation pipeline [default: bgc]
--resume PKL Resume training [default: noresume]
--freezed INT Freeze-D [default: 0 layers]
--fp32 BOOL Disable mixed-precision training
--nhwc BOOL Use NHWC memory format with FP16
--nobench BOOL Disable cuDNN benchmarking
--allow-tf32 BOOL Allow PyTorch to use TF32 internally
--workers INT Override number of DataLoader workers
--beta0 FLOAT Beta_0
--help Show this message and exit.
Generate 50,000 images using the model weights associated with the lowest Fréchet Inception Distance (FID)5 attained during training.
Images can be generated by providing --network and --outdir arguments to generator.sh.
./bash_scripts/generator.sh
Usage: generate.py [OPTIONS]
Generate images using pretrained network pickle.
Examples:
# Generate curated MetFaces images without truncation (Fig.10 left)
python generate.py --outdir=out --trunc=1 --seeds=85,265,297,849 \
--network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metfaces.pkl
# Generate uncurated MetFaces images with truncation (Fig.12 upper left)
python generate.py --outdir=out --trunc=0.7 --seeds=600-605 \
--network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metfaces.pkl
# Generate class conditional CIFAR-10 images (Fig.17 left, Car)
python generate.py --outdir=out --seeds=0-35 --class=1 \
--network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/cifar10.pkl
# Render an image from projected W
python generate.py --outdir=out --projected_w=projected_w.npz \
--network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metfaces.pkl
Options:
--network TEXT Network pickle filename [required]
--seeds NUM_RANGE List of random seeds
--trunc FLOAT Truncation psi [default: 1]
--class INTEGER Class label (unconditional if not specified)
--noise-mode [const|random|none]
Noise mode [default: const]
--projected-w FILE Projection result file
--outdir DIR Where to save the output images [required]
--help Show this message and exit.
find_best_fid.sh given the --fname argument can be used to determine which weights were associated with the lowest FID score.
./bash_scripts/find_best_fid.sh
usage: find_best_fid.py [-h] [-f FNAME]
Required Arguments:
-f FNAME, --fname FNAME
Path to the StyleGAN2-ADA output JSON file with metric information, i.e.
the "metric-fid50k_full" JSON file.
Evaluate FID and Fréchet SwAV Distance (FSD)6 with the StudioGAN3 by providing --dset1, --dset2, --eval_backbone (either InceptionV3_torch7 or SwAV_torch8), and --out_path to the eval_fd.sh script.
The folder names within the --dset1 and --dset2 directories that contain the images to be evaluated must match (such as class0).
The batch size --batch_size argument (defaults to 64) can be updated if memory issues are encountered.
eval_fd.sh uses the studiogan Docker container.
If you have problems with the compatibility of the PyTorch version with your CUDA installation, you can either use the alex4727/experiment:pytorch113_cuda116 container or change the PyTorch installation command within the PyTorch-StudioGAN/Dockerfile and rebuild the studiogan container.
When evaluating a generative distribution, the first dataset consists of real images (training images), and the second consists of the generated images (or vice versa). When determining a baseline, the first and second datasets come from a random split of the real images.
./bash_scripts/eval_fd.sh
usage: evaluate.py [-h] [-metrics EVAL_METRICS [EVAL_METRICS ...]] [--post_resizer POST_RESIZER] [--eval_backbone EVAL_BACKBONE] [--dset1 DSET1]
[--dset1_feats DSET1_FEATS] [--dset1_moments DSET1_MOMENTS] [--dset2 DSET2] [--batch_size BATCH_SIZE] [--seed SEED] [-DDP]
[--backend BACKEND] [-tn TOTAL_NODES] [-cn CURRENT_NODE] [--num_workers NUM_WORKERS] --out_path OUT_PATH
optional arguments:
-h, --help show this help message and exit
-metrics EVAL_METRICS [EVAL_METRICS ...], --eval_metrics EVAL_METRICS [EVAL_METRICS ...]
evaluation metrics to use during training, a subset list of ['fid', 'is', 'prdc'] or none
--post_resizer POST_RESIZER
which resizer will you use to evaluate GANs in ['legacy', 'clean', 'friendly']
--eval_backbone EVAL_BACKBONE
[InceptionV3_tf, InceptionV3_torch, ResNet50_torch, SwAV_torch, DINO_torch, Swin-T_torch]
--dset1 DSET1 specify the directory of the folder that contains dset1 images (real).
--dset1_feats DSET1_FEATS
specify the path of *.npy that contains features of dset1 (real). If not specified, StudioGAN will automatically extract
feat1 using the whole dset1.
--dset1_moments DSET1_MOMENTS
specify the path of *.npy that contains moments (mu, sigma) of dset1 (real). If not specified, StudioGAN will
automatically extract moments using the whole dset1.
--dset2 DSET2 specify the directory of the folder that contains dset2 images (fake).
--batch_size BATCH_SIZE
batch_size for evaluation
--seed SEED seed for generating random numbers
-DDP, --distributed_data_parallel
--backend BACKEND cuda backend for DDP training \in ['nccl', 'gloo']
-tn TOTAL_NODES, --total_nodes TOTAL_NODES
total number of nodes for training
-cn CURRENT_NODE, --current_node CURRENT_NODE
rank of the current node
--num_workers NUM_WORKERS
--out_path OUT_PATH output file to put metrics into
The Fréchet Radiomics Distance (FRD)4 can be calculated by providing the paths to the two directories containing the datasets to be evaluated to eval_frd.sh.
./bash_scripts/eval_frd.sh
This code uses a fork of the official StyleGAN2-ADA repository to reconstruct the images with a trained StyleGAN2-ADA model via backpropagation.
You'll need to be on the proj_dir branch of the repository to use our expanded capabilities of reconstructing all images in a given directory and saving images in grayscale.
To reconstruct the images, provide the --network, --target, and --outdir arguments to the projector.sh script.
The --mode optional argument can be changed to save images in color (defaults to grayscale).
./bash_scripts/projector.sh
Usage: projector.py [OPTIONS]
Project given image to the latent space of pretrained network pickle.
Examples:
python projector.py --outdir=out --target=~/mytargetimg.png \
--network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/ffhq.pkl
Options:
--network TEXT Network pickle filename [required]
--target FILE Directory with the target image file to project to
[required]
--num-steps INTEGER Number of optimization steps [default: 1000]
--seed INTEGER Random seed [default: 303]
--save-video BOOLEAN Save an mp4 video of optimization progress [default:
True]
--outdir DIR Where to save the output images [required]
--mode TEXT Whether to reconstruct color (RGB) or grayscale (L)
images [default: RGB]
--help Show this message and exit.
The distances between images and their reconstructions can be calculated by providing the --orig_dir, --recon_dir, and --out_path arguments to eval_recon.sh.
--distance controls whether the mean-squared error or the Wasserstein distance is used.
--region controls whether the distance is calculated within a full image, within the human body, or within a patch.
If patch is specified, then --patch_size determines the patch size and --patch_path provides a path to a directory where the patch with the highest distance will be saved, along with its reconstruction, which is placed to the right of the original patch.
./scripts/eval_recon.sh
usage: eval_recon.py [-h] [-o ORIG_DIR] [-op OUT_PATH] [-r RECON_DIR] [-d DISTANCE] [-ps PATCH_SIZE] [-pp PATCH_PATH] [-reg REGION]
Required Arguments:
-o ORIG_DIR, --orig_dir ORIG_DIR
Path to the directory containing the original images.
-op OUT_PATH, --out_path OUT_PATH
Path to csv file to write distances to.
-r RECON_DIR, --recon_dir RECON_DIR
Path to the directory containing the reconstructed images.
Optional Arguments:
-d DISTANCE, --distance DISTANCE
One of [MSE, WD] for mean-squared error and Wasserstein distance. Defaults to MSE.
-ps PATCH_SIZE, --patch_size PATCH_SIZE
Size of the patches when the region argument is "patch". Default: 32.
-pp PATCH_PATH, --patch_path PATCH_PATH
Folder to save the highest-scoring patches and their reconstructions to. Default to None. If None, no patches will be saved.
-reg REGION, --region REGION
The region of the image that the reconstruction metric should be calculated on. Options: ["full","body","patch"]. Default: "full".
The baseline test dataset and each anomaly test dataset must be placed in separate directories prior to the distance calculations. Distances must be calculated on each test directory, resulting in a CSV file of distances for each test dataset.
Areas under the receiver operating characteristic curve (AUROCs) can be calculated by providing the paths to the CSV files containing the distances for the baseline --baseline and anomaly --anomaly test datasets (i.e., the outputs of the previous section) to eval_ad.sh, along with a path specifying where to put the CSV with the calculated AUROCs.
By default, bootstrapping with 50 samples will be performed.
The number of bootstrap samples can be changed with --num_samp.
The final CSV file will contain the AUROCs for all bootstraps.
The script prints out the mean and standard deviation of these AUROCs.
./bash_scripts/eval_ad.sh
usage: eval_ad.py [-h] -a ANOMALY -b BASELINE -o OUT_PATH [-n NUM_SAMP]
Required Arguments:
-a ANOMALY, --anomaly ANOMALY
Path to csv file containing the distances for the anomalous dataset.
-b BASELINE, --baseline BASELINE
Path to csv file containing the distances for the baseline data.
-o OUT_PATH, --out_path OUT_PATH
Path to csv file to write AUROCS to.
Optional Arguments:
-n NUM_SAMP, --num_samp NUM_SAMP
The number of bootstrap samples to use.
Permutation tests can be performed with permutation_test.sh by providing paths to the CSV files with bootstrapped AUROCs to be compared --csv1 and --csv1.
This is a one-sided permutation test that evaluates whether the mean of the first population is larger than the mean of the second.
The simulated power of each calculation will be printed, along with the p-value of the executed test.
By default, 100,000 permutations will be performed for each test.
For the simulated power, the default significance level is
./bash_scripts/permutation_test.sh
usage: permutation_test.py [-h] -c1 CSV1 -c2 CSV2 [-n1 COL_NAME1] [-n2 COL_NAME2]
Required Arguments:
-c1 CSV1, --csv1 CSV1
Path to CSV file containing the first population.
-c2 CSV2, --csv2 CSV2
Path to CSV file containing the second population.
Optional Arguments:
-n1 COL_NAME1, --col_name1 COL_NAME1
Name of column containing the first population. Defaults to AUROC.
-n2 COL_NAME2, --col_name2 COL_NAME2
Name of column containing the second population. Defaults to AUROC.
The data from The University of Texas MD Anderson Cancer Center may be made available upon request, in compliance with institutional review board requirements.
Data can be windowed with window_level.sh by providing the directory containing the images to be windowed in NIfTI format --in_dir and the output directory --out_dir.
Window width and level can be changed from the defaults of 350 and 50 with --window_width and --window_level.
./bash_scripts/window_level.sh
usage: window_level.py [-h] -i IN_DIR -o OUT_DIR [-w WINDOW_WIDTH] [-l WINDOW_LEVEL]
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing images to be windowed. Images must be in NIfTI format.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put the windowed images into.
Optional Arguments:
-w WINDOW_WIDTH, --window_width WINDOW_WIDTH
Window width. Defaults to 350.
-l WINDOW_LEVEL, --window_level WINDOW_LEVEL
Window level. Defaults to 50.
All images must be in NIfTI format before windowing.
For convenience, we have provided dicom_2_nifti.sh to convert DICOM files to NIfTI.
./bash_scripts/dicom_2_nifti.sh
usage: dicom_2_nifti.py [-h] -i IN_DIR -o OUT_DIR
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing DICOMs to be converted.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put the NIfTIs.
Three-dimensional images can be converted to two-dimensional axial slices by providing --in_dir and --out_dir to slice.sh.
If --mask_dir is provided, only slices that contain the given structure will be saved.
The masks must have the same filename as the original image.
The mask filename must not contain a period.
The slices containing the mask must be consecutive.
./bash_scripts/slice.sh
usage: slice.py [-h] -i IN_DIR -o OUT_DIR [-m MASK_DIR]
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing the 3D NIfTI images to be sliced.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put the 2D NIfTI slices.
Optional Arguments:
-m MASK_DIR, --mask_dir MASK_DIR
Path to the directory containing masks. The masks must have the same filename as the original image. The mask filename must not
contain a period. The slices containing the mask must be consecutive. All saved slices will contain the given structure. If None,
all slices will be saved. Default is None.
NIfTI slices can be converted to PNG images with png_conversion.sh.
./bash_scripts/png_conversion.sh
usage: png_conversion.py [-h] -i IN_DIR -o OUT_DIR
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing NIfTI slices to be converted.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put the PNG images.
The ChestX-ray149 dataset is downloadable from Box.
Images can be rescaled to 512x512 with rescale_images.sh by providing the paths to the directory containing the original images --in_dir and the directory to put the resized images into --out_dir.
./bash_scripts/rescale_images.sh
usage: rescale_images.py [-h] -i IN_DIR -o OUT_DIR [-r RESOLUTION]
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing images to be resized.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put resized images into.
Optional Arguments:
-r RESOLUTION, --resolution RESOLUTION
Desired output resolution. Defaults to 512.
All available chest radiographs were downloaded from the Medical Imaging and Data Resource Center (MIDRC) in November 2022. Specifically, under "Body Part Examined", "CHEST" and "PORT CHEST" were selected and "DX" and "CR" were selected under "Study Modality". The exact Case IDs, Study UIDs, and filenames associated with the 64,373 radiographs downloaded are available via request for replication purposes, in compliance with MIDRC policies.
The preprocess_midrc script can be used to perform the preprocessing: slice, resize to 512x512, rescale to [0, 255], convert to unsigned 8-bit integers, and write to PNG.
The DICOMs within --in_dir should be organized first by the body part examined and then by the modality.
./bash_scripts/preprocess_midrc.sh
usage: preprocess_midrc.py [-h] -i IN_DIR -o OUT_DIR
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to directory that contains the original DICOMs in subdirectories. The first subset of
directories should be named after the body part examined. The second subset should be named
after the modality.
-o OUT_DIR, --out_dir OUT_DIR
Path to directory to put PNGs into.
Gaussian noise and blur can be added to images with add_noise.sh.
./bash_scripts/add_noise.sh
usage: add_noise.py [-h] -i IN_DIR -o OUT_DIR [-k KERNEL_SIZE] [-m MEAN] [-s SIGMA] [-t TYPE] [-v VAR]
Required Arguments:
-i IN_DIR, --in_dir IN_DIR
Path to the directory containing images to be manipulated.
-o OUT_DIR, --out_dir OUT_DIR
Path to the directory to put the manipulated images into.
Optional Arguments:
-k KERNEL_SIZE, --kernel_size KERNEL_SIZE
Size of Gaussian kernel (blur). Defaults to (5,5).
-m MEAN, --mean MEAN Mean of Gaussian distribution (noise). Defaults to 0.
-s SIGMA, --sigma SIGMA
Standard deviation of kernel (blur). Defaults to 0.
-t TYPE, --type TYPE Type of image manipulation: Gaussian noise (n) or blur (b). Defaults to n.
-v VAR, --var VAR Variance of Gaussian distribution (noise). Defaults to 0.01.
Datasets can be randomly split in half with split_dataset.sh.
./bash_scripts/split_dataset.sh
usage: split_dataset.py [-h] --in_dir IN_DIR --out_dir1 OUT_DIR1 --out_dir2 OUT_DIR2
Required Arguments:
--in_dir IN_DIR Path to folder containing dataset to split.
--out_dir1 OUT_DIR1 Path to folder to put first half of the dataset into.
--out_dir2 OUT_DIR2 Path to folder to put second half of the dataset into.
The ChestX-ray14 StyleGAN2-ADA weights that were used for anomaly detection within radiographs from MIDRC have been published on Zenodo.
Research reported in this publication was supported in part by resources of the Image Guided Cancer Therapy Research Program at The University of Texas MD Anderson Cancer Center; by the National Institutes of Health/National Cancer Institute under award numbers P30CA016672, R01CA235564, and R01CA221971; by the Diagnostic Imaging – Summer Training and Experiences Partnership (DI-STEP); by the Tumor Measurement Initiative through the MD Anderson Strategic Initiative Development Program (STRIDE); by the Helen Black Image Guided Fund; and by a generous gift from the Apache Corporation.
The imaging and associated clinical data downloaded from MIDRC (The Medical Imaging and Data Resource Center) and used for research in this publication were made possible by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) of the National Institutes of Health under contract 75N92020D00021. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
If you have found our work useful, we would appreciate a citation to our extended abstract.
@misc{woodland2023stylegan2basedoutofdistributiondetectionmedical,
title={StyleGAN2-based Out-of-Distribution Detection for Medical Imaging},
author={McKell Woodland and John Wood and Caleb O'Connor and Ankit B. Patel and Kristy K. Brock},
year={2023},
eprint={2307.10193},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2307.10193},
}
- McKell Woodland et al. StyleGAN2-based out-of-distribution detection for medical imaging. arXiv:2307.10193. Accepted abstract.
- Tero Karras et al. Training generative adversarial networks with limited data. In NeurIPS 2020; Curran Associates, Inc; 33:12104-12114; 2020. Proceedings.
- Minguk Kang et al. StudioGAN: A taxonomy and benchmark of GANs for image synthesis. TPAMI, 45(12):15725-15742. Manuscript.
- Osuala et al. Towards learning contrast kinetics with multi-condition latent diffusion models. In MICCAI 2024; Springer, Cham; 15005:713-723; 2024. Proceedings.
- Martin Heusel et al. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NIPS 2017; Curran Associates, Inc.; 30:6629-6640; 2017. Proceedings.
- Stanislav Morozov et al. On self-supervised image representations for GAN evaluation. In ICLR 2021; 2021. OpenReview
- Christian Szegedy et al. Going deeper with convolutions. In CVPR 2015, IEEE, pp. 1-9, 2015.Proceedings.
- Mathilde Caron et al. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS 2020; Curran Associates, Inc.; 33:9912-9924; 2020. Proceedings.
- Xiaosong Wang et al. ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR 2017; IEEE; 3462-3471; 2017. Proceedings.