krekiehn/aucmedi

a framework for Automated Classification of Medical Images

The open-source software AUCMEDI allows fast setup of medical image classification pipelines with state-of-the-art methods via an intuitive, high-level Python API or via an AutoML deployment through Docker/CLI.

AUCMEDI provides several core features:

• Wide range of 2D/3D data entry options with interfaces to the most common medical image formats such as DICOM, MetaImage, NifTI, PNG or TIF already supplied.
• Selection of pre-processing methods for preparing images, such as augmentation processes, color conversions, windowing, filtering, resizing and normalization.
• Use of deep neural networks for binary, multi-class as well as multi-label classification and efficient methods against class imbalances using modern loss functions such as focal loss.
• Library from modern architectures, like ResNet up to EfficientNet and Vision-Transformers (ViT)⁠.
• Complex ensemble learning techniques (combination of predictions) using test-time augmentation, bagging via cross-validation or stacking via logistic regressions.
• Explainable AI to explain opaque decision-making processes of the models using activation maps such as Grad-CAM or backpropagation.
• Automated Machine Learning (AutoML) mentality to ensure easy deployment, integration and maintenance of complex medical image classification pipelines (Docker).

Resources

• Website: AUCMEDI Website - Home
• Git Repository: GitHub - frankkramer-lab/aucmedi
• Documentation: AUCMEDI Wiki - API Reference
• Getting Started: AUCMEDI Website - Getting Started
• Examples: AUCMEDI Wiki - Examples
• Tutorials: AUCMEDI Wiki - Tutorials
• Applications: AUCMEDI Wiki - Applications
• PyPI Package: PyPI - aucmedi
• Docker Image: GitHub - ghcr.io/frankkramer-lab/aucmedi
• Zenodo Repository: Zenodo - AUCMEDI

Getting started: 60 seconds to automated medical image classification

Simply install AUCMEDI with a single line of code via pip.

Install AUCMEDI via PyPI

pip install aucmedi

Now, you can build a state-of-the-art medical image classification pipeline via the standardized AutoML interface or a custom pipeline with the framework interface.

AutoML

Train a model and classify unknown images

# Run training with default arguments, but a specific architecture
aucmedi training --architecture "DenseNet121"

# Run prediction with default arguments
aucmedi prediction

Framework

Your custom pipeline with just the 3 AUCMEDI pillars:

• Pillar #1: input_interface() for obtaining general dataset information
• Pillar #2: NeuralNetwork() for the deep learning model
• Pillar #3: DataGenerator() for a powerful interface to load any images/volumes into your model

Build a pipeline

# AUCMEDI library
from aucmedi import *

# Pillar #1: Initialize input data reader
ds = input_interface(interface="csv",
                     path_imagedir="/home/muellerdo/COVdataset/ct_slides/",
                     path_data="/home/muellerdo/COVdataset/classes.csv",
                     ohe=False,           # OHE short for one-hot encoding
                     col_sample="ID", col_class="PCRpositive")
(index_list, class_ohe, nclasses, class_names, image_format) = ds

# Pillar #2: Initialize a DenseNet121 model with ImageNet weights
model = NeuralNetwork(n_labels=nclasses, channels=3,
                       architecture="2D.DenseNet121",
                       pretrained_weights=True)

Train a model and use it!

# Pillar #3: Initialize training Data Generator for first 1000 samples
train_gen = DataGenerator(samples=index_list[:1000],
                          path_imagedir="/home/muellerdo/COVdataset/ct_slides/",
                          labels=class_ohe[:1000],
                          image_format=image_format,
                          resize=model.meta_input,
                          standardize_mode=model.meta_standardize)
# Run model training with Transfer Learning
model.train(train_gen, epochs=20, transfer_learning=True)

# Pillar #3: Initialize testing Data Generator for 500 samples
test_gen = DataGenerator(samples=index_list[1000:1500],
                         path_imagedir="/home/muellerdo/COVdataset/ct_slides/",
                         labels=None,
                         image_format=image_format,
                         resize=model.meta_input,
                         standardize_mode=model.meta_standardize)
# Run model inference for unknown samples
preds = model.predict(test_gen)

# preds <-> NumPy array with shape (500,2)
# -> 500 predictions with softmax probabilities for our 2 classes

How to cite / More information

AUCMEDI is currently unpublished. But coming soon!

In the meantime:
Please cite our application manuscript as well as the AUCMEDI GitHub repository:

Mayer, S., Müller, D., & Kramer F. (2022). Standardized Medical Image Classification
across Medical Disciplines. [Preprint] https://arxiv.org/abs/2210.11091.

@article{AUCMEDIapplicationMUELLER2022,
  title={Standardized Medical Image Classification across Medical Disciplines},
  author={Simone Mayer, Dominik Müller, Frank Kramer},
  year={2022}
  eprint={2210.11091},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Müller, D., Mayer, S., Hartmann, D., Schneider, P., Soto-Rey, I., & Kramer, F. (2022).
AUCMEDI: a framework for Automated Classification of Medical Images (Version X.Y.Z) [Computer software].
https://doi.org/10.5281/zenodo.6633540. GitHub repository. https://github.com/frankkramer-lab/aucmedi

Thank you for citing our work.

Lead Author

Dominik Müller
Email: dominik.mueller@informatik.uni-augsburg.de
IT-Infrastructure for Translational Medical Research
University Augsburg
Bavaria, Germany

License

This project is licensed under the GNU GENERAL PUBLIC LICENSE Version 3.
See the LICENSE.md file for license rights and limitations.