Reinventing 2D Convolutions for 3D Images (arXiv)
IEEE Journal of Biomedical and Health Informatics (IEEE JBHI), 2021 (DOI)
News:
- 2022.01.26 - ACS ConvNeXt supported.
- 2021.12.17 - torch 1.10 supported & pip installation supported.
- 2021.4.19 - torch 1.8 supported.
- ACS convolution aims at a plug-and-play replacement of standard 3D convolution, for 3D medical images.
- ACS convolution enables 2D-to-3D transfer learning, which consistently provides significant performance boost in our experiments.
- Even without pretraining, ACS convolution is comparable to or even better than 3D convolution, with smaller model size and less computation.
If you want to use this class, you have two options:
A) Install ACSConv as a standard Python package from PyPI:
pip install ACSConv
Alternatively, via conda:
conda install acsconv -c conda-forge
B) Simply copy and paste it in your project;
You could run the test.py
to validate the installation. (If you want to test the validity of pip installation, please move this test.py
file outside of this git project directory, otherwise it is testing the code inside the project instead of pip installation.)
Recommended PyTorch versions
torch>=1.0.0 and torch<=1.10.0
Compatibility of other PyTorch versions are not guaranteed. They should be compatible if the relevant APIs are consistent. Feel free to try.
All libraries needed to run the included experiments (base requirements included).
fire
jupyterlab
matplotlib
pandas
tqdm
sklearn
tensorboardx
acsconv
the core implementation of ACS convolution, including the operators, models, and 2D-to-3D/ACS model converters.operators
: include ACSConv, SoftACSConv and Conv2_5d.converters
: include converters which convert 2D models to 3d/ACS/Conv2_5d counterparts.models
: Native ACS models.
experiments
the scripts to run experiments.mylib
: the lib for running the experiments.poc
: the scripts to run proof-of-concept experiments.lidc
: the scripts to run LIDC-IDRI experiments.
import torch
from torchvision.models import resnet18
from acsconv.converters import ACSConverter
# model_2d is a standard pytorch 2D model
model_2d = resnet18(pretrained=True)
B, C_in, H, W = (1, 3, 64, 64)
input_2d = torch.rand(B, C_in, H, W)
output_2d = model_2d(input_2d)
model_3d = ACSConverter(model_2d)
# once converted, model_3d is using ACSConv and capable of processing 3D volumes.
B, C_in, D, H, W = (1, 3, 64, 64, 64)
input_3d = torch.rand(B, C_in, D, H, W)
output_3d = model_3d(input_3d)
import torch
from acsconv.operators import ACSConv, SoftACSConv
B, C_in, D, H, W = (1, 3, 64, 64, 64)
x = torch.rand(B, C_in, D, H, W)
# ACSConv to process 3D volumnes
conv = ACSConv(in_channels=3, out_channels=10, kernel_size=3, padding=1)
out = conv(x)
# SoftACSConv to process 3D volumnes
conv = SoftACSConv(in_channels=3, out_channels=10, kernel_size=3, padding=1)
out = conv(x)
import torch
from acsconv.models.acsunet import ACSUNet
unet_3d = ACSUNet(num_classes=3)
B, C_in, D, H, W = (1, 1, 64, 64, 64)
input_3d = torch.rand(B, C_in, D, H, W)
output_3d = unet_3d(input_3d)