Self-supervised learning algorithms provide a way to train Deep Neural Networks in an unsupervised way using contrastive losses. The idea is to learn a representation which can discriminate between negative examples and be as close as possible to augmentations and transformations of itself. In this approach, we first train a ResNet on the unlabeled dataset which is then fine-tuned on a relatively small labeled one. This approach drastically reduces the amount of labeled data required, a big problem in applying deep learning in the real world. Surprisingly, this approach actually leads to increase in robustness as well as raw performance, when compared to fully supervised counterparts, even with the same architecture.
In case, the user wants to skip the pre-training part, the pre-trained weights can be downloaded from here to use for fine-tuning tasks and directly skip to the second part of the tutorial which is using the 'ssl_finetune_train.py'.
1.) Download the two datasets TCIA-Covid19
& BTCV (More detail about them in the Data section)
2.) Modify the paths for data_root, json_path & logdir in ssl_script_train.py
3.) Run the 'ssl_script_train.py'
4.) Modify the paths for data_root, json_path, pre-trained_weights_path from 2.) and
logdir_path in 'ssl_finetuning_train.py'
5.) Run the 'ssl_finetuning_script.py'
6.) And that's all folks, use the model to your needs
Pre-training Dataset: The TCIA Covid-19 dataset was used for generating the pre-trained weights. The dataset contains a total of 771 3D CT Volumes. The volumes were split into training and validation sets of 600 and 171 3D volumes correspondingly. The data is available for download at this link. If this dataset is being used in your work, please use [1] as reference. A json file is provided which contains the training and validation splits that were used for the training. The json file can be found in the json_files directory of the self-supervised training tutorial.
Fine-tuning Dataset: The dataset from Beyond the Cranial Vault Challenge (BTCV) 2015 hosted at MICCAI, was used as a fully supervised fine-tuning task on the pre-trained weights. The dataset consists of 30 3D Volumes with annotated labels of up to 13 different organs [2]. There are 3 json files provided in the json_files directory for the dataset. They correspond to having different number of training volumes ranging from 6, 12 and 24. All 3 json files have the same validation split.
References:
1.) Harmon, Stephanie A., et al. "Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets." Nature communications 11.1 (2020): 1-7.
2.) Tang, Yucheng, et al. "High-resolution 3D abdominal segmentation with random patch network fusion." Medical Image Analysis 69 (2021): 101894.
For pre-training a modified version of ViT [1] has been used, it can be referred here from MONAI. The original ViT was modified by attachment of two 3D Convolutional Transpose Layers to achieve a similar reconstruction size as that of the input image. The ViT is the backbone for the UNETR [2] network architecture which was used for the fine-tuning fully supervised tasks.
The pre-trained backbone of ViT weights were loaded to UNETR and the decoder head still relies on random initialization for adaptability of the new downstream task. This flexibility also allows the user to adapt the ViT backbone to their own custom created network architectures as well.
References:
1.) Dosovitskiy, Alexey, et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXiv preprint arXiv:2010.11929 (2020).
2.) Hatamizadeh, Ali, et al. "Unetr: Transformers for 3d medical image segmentation." arXiv preprint arXiv:2103.10504 (2021).
The pre-training pipeline has two aspects to it (Refer figure shown below). First, it uses augmentation (top row) to mutate the data and second, it utilizes regularized contrastive loss [3] to learn feature representations of the unlabeled data. The multiple augmentations are applied on a randomly selected 3D foreground patch from a 3D volume. Two augmented views of the same 3D patch are generated for the contrastive loss as it functions by drawing the two augmented views closer to each other if the views are generated from the same patch, if not then it tries to maximize the disagreement. The CL offers this functionality on a mini-batch.
The augmentations mutate the 3D patch in various ways, the primary task of the network is to reconstruct the original image. The different augmentations used are classical techniques such as in-painting [1], out-painting [1] and noise augmentation to the image by local pixel shuffling [2]. The secondary task of the network is to simultaneously reconstruct the two augmented views as similar to each other as possible via regularized contrastive loss [3] as its objective is to maximize the agreement. The term regularized has been used here because contrastive loss is adjusted by the reconstruction loss as a dynamic weight itself.
The below example image depicts the usage of the augmentation pipeline where two augmented views are drawn of the same 3D patch:
Multiple axial slices of a 96x96x96 patch are shown before the augmentation (Ref Original Patch in the above figure).
Augmented View 1 & 2 are different augmentations generated via the transforms on the same cubic patch. The objective
of the SSL network is to reconstruct the original top row image from the first view. The contrastive loss
is driven by maximizing agreement of the reconstruction based on input of the two augmented views.
matshow3d from monai.visualize was used for creating this figure, a tutorial for using can be found here
References:
1.) Pathak, Deepak, et al. "Context encoders: Feature learning by inpainting." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
2.) Chen, Liang, et al. "Self-supervised learning for medical image analysis using image context restoration." Medical image analysis 58 (2019): 101539.
3.) Chen, Ting, et al. "A simple framework for contrastive learning of visual representations." International conference on machine learning. PMLR, 2020.
Training Hyper-Parameters for SSL:
Epochs: 300
Validation Frequency: 2
Learning Rate: 1e-4
Batch size: 4 3D Volumes (Total of 8 as 2 samples were drawn per 3D Volume)
Loss Function: L1
Contrastive Loss Temperature: 0.005
Training Hyper-parameters for Fine-tuning BTCV task (All settings have been kept consistent with prior
UNETR 3D
Segmentation tutorial):
Number of Steps: 30000
Validation Frequency: 100 steps
Batch Size: 1 3D Volume (4 samples are drawn per 3D volume)
Learning Rate: 1e-4
Loss Function: DiceCELoss
L1 error reported for training and validation when performing the SSL training. Please note contrastive loss is not L1.
| Training Volumes | Validation Volumes | Random Init Dice score | Pre-trained Dice Score | Relative Performance Improvement |
|---|---|---|---|---|
| 6 | 6 | 63.07 | 70.09 | ~11.13% |
| 12 | 6 | 76.06 | 79.55 | ~4.58% |
| 24 | 6 | 78.91 | 82.30 | ~4.29% |
@article{Arijit Das,
title={Self-supervised learning for medical data},
author={Arijit Das},
journal={https://github.com/das-projects/selfsupervised-learning},
year={2020}
}