MedMNIST Experiments

Training and evaluation scripts to reproduce both 2D and 3D experiments in our MedMNIST paper, including PyTorch, auto-sklearn, AutoKeras and Google AutoML Vision together with their weights ;)

Code Structure

• MedMNIST2D/: training and evaluation scripts of MedMNIST2D

  • models.py: ResNet-18 and ResNet-50 models (for small-image datasets like CIFAR-10/100)
  • train_and_eval_pytorch.py: training and evaluation script implemented with PyTorch
  • train_and_eval_autokeras.py: training and evaluation script of Autokeras
  • train_and_eval_autosklearn.py: training and evaluation script of auto-sklearn
  • eval_google_automl_vision.py: evaluation script of models trained by Google AutoML Vision

• MedMNIST3D/: training and evaluation scripts of MedMNIST3D

  • models.py: ResNet-18 and ResNet-50 models (for small-image datasets like CIFAR-10/100), basically same as MedMNIST2D/models.py
  • train_and_eval_pytorch.py: training and evaluation script implemented with PyTorch
  • train_and_eval_autokeras.py: training and evaluation script of Autokeras
  • train_and_eval_autosklearn.py: training and evaluation script of auto-sklearn

Installation and Requirements

This repository is working with MedMNIST official code and PyTorch.

1. Setup the required environments and install medmnist as a standard Python package:

    pip install medmnist
   

2. Check whether you have installed the latest version:

    >>> import medmnist
    >>> print(medmnist.__version__)
   

3. The code requires common Python environments for machine learning; Basically, it was tested with

• Python 3 (Anaconda 3.6.3 specifically)

• PyTorch==1.3.1

• autokeras==1.0.15

• auto-sklearn==0.10.0

• tensorflow==2.3.0

  Higher (or lower) versions should also work (perhaps with minor modifications).

4. For MedMNIST3D, our code additionally requires ACSConv. Install it through pip via the command bellow:

    pip install git+git://github.com/M3DV/ACSConv.git
   

   Then you can check use the cool 2.5D, 3D and ACS model convertor as follows:

   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)

5. Download the model weights and predictions from Zenodo.

   • weights_*.zip:
     • PyTorch, AutoKeras and Google AutoML Vision are provided for MedMNIST2D.
     • PyTorch and AutoKeras are provided for MedMNIST3D.
     • If you are using PyTorch model weights, please note that the ResNet18_224 / ResNet50_224 models are trained with images resized to 224 x 224 by PIL.Image.NEAREST.
     • Snapshots for auto-sklearn are not uploaded due to the embarrassingly large model sizes (lots of model ensemble).
   • predictions.zip: We also provide all standard prediction files by PyTorch, auto-sklearn, AutoKeras and Google AutoML Vision, which works with medmnist.Evaluator. Each file is named as {flag}_{split}_[AUC]{auc:.3f}_[ACC]{acc:.3f}@{run}.csv, e.g., bloodmnist_test_[AUC]0.997_[ACC]0.957@autokeras_3.csv.