Skip to content

In this work we present a task-agnostic Multimodal Variational Aversarial Active Learning (M-VAAL) for sampling the most informative samples for annotation in various Medical Image Analysis Downstream tasks, such as segmentation, and classification.

License

Notifications You must be signed in to change notification settings

Bidur-Khanal/MVAAL-medical-images

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MVAAL-medical-images

In this work, we present a task-agnostic Multimodal Variational Aversarial Active Learning (M-VAAL) for sampling the most informative samples for annotation in various Medical Image Analysis Downstream tasks, such as segmentation, and classification.

Active Learning Pipeline

Our method modified task-agnostic VAAL framework to enable the use of multi-modal information in the active learning sampling phase. pipeline

Installation

The required Python3 packages can be installed using

conda config --add channels conda-forge
conda create -n <environment-name> --file requirements.txt

Datasets

We experimented with 210 High-Grade Gliomas BraTS2018 and COVID-QU-Ex (COVID-19 infection segmentation) datasets. Data preprocessing steps are shown in BRATS_dataset_creator.ipynb and covid_visualizer.ipynb

Experiments

To run the code use:

python3 main.py --dataset <Dataset> --expt <Experiment Name> --method <Sampling Method> --task_type <Downstream task, e.g: segmentation>

In our paper, we used the following combination of arguments:

--dataset brats_MIUA_HGG_3channel_classification --task_type multi_label_classification
OR 
--dataset brats_MIUA_HGG_3channel --task_type segmentation
OR
--dataset COVID_dataset_3_classes --task_type multi_class_classification

There are several hyperparameters that need to properly tunned to optimize the performance of M-VAAL and VAAL. These parameters are sensitive to dataset and task type.

--adversary_param --latent_dim --beta --num_vae_steps and --num_adv_steps --mse_gamma2 --alpha1 --alpha2

Results

We compared M-VAAL against VAAL and randomsampling three downstream tasks: tumor segmentation, multi-label tumor type classification, and multi-class chest-Xray Infection classifcation.


BraTS Whole Tumor Segmentation


BraTS Tumor Type Multi-Label Classification


Chest X-ray Multi-Class classification

Qualitative Results


White denotes the region overlapped between two segmentation maps, blue denotes the region missed by the first method compared to second, and red denotes the region added by the first method.

This work has been accepted at 27th Conference on Medical Image Understanding and Analysis 2023. Please use the following bibtex to cite our paper.

@article{khanal2023m,
  title={M-VAAL: Multimodal Variational Adversarial Active Learning for Downstream Medical Image Analysis Tasks},
  author={Khanal, Bidur and Bhattarai, Binod and Khanal, Bishesh and Stoyanov, Danail and Linte, Cristian A},
  journal={arXiv preprint arXiv:2306.12376},
  year={2023}
}

About

In this work we present a task-agnostic Multimodal Variational Aversarial Active Learning (M-VAAL) for sampling the most informative samples for annotation in various Medical Image Analysis Downstream tasks, such as segmentation, and classification.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published