To simultaneously contend with the lack of raw k-space data for learning accelerated MRI reconstruction and its sensitivity to distribution shifts, we have developed a consistency-based semi-supervised MRI reconstruction paradigm termed VORTEX. In this formulation, for datasets where we only have undersampled MRI data, we simulate distribution shifts using our knowledge of the MRI forward model and MRI physics. To build invariance to these distribution shifts (such as motion or noise), we enforce consistency between the augmented and unaugmented versions of the undersampled images. This technique improves performance on in-distribution non-perturbed data, as well as out-of-distribution perturbed data.

As an extreme example for data-limited MRI reconstruction, we consider the scenario where no training data is available. Here, we use the Deep Decoder framework, which is an extension of the Deep Image Prior. By using the fundamental architecture of a convolutional neural network as an image regularizer, we can performed accelerated MRI reconstruction without any training data, with image quality on par with supervised paradigms.

Select Publications:

• Desai A, Ozturkler B, Sandino C, Hargreaves B, Pauly J, and Chaudhari A. Noise2Recon: A Semi-Supervised Framework for Joint MRI Reconstruction and Denoising using Limited Data. Intl Soc Magn Reson Med, (virtual), 2021.
• Van Veen D, Desai A, Heckel R, and Chaudhari A. Using Untrained Convolutional Neural Networks to Accelerate MRI in 2D and 3D. Intl Soc Magn Reson Med, (virtual), 2021.
• VORTEX.
• Darestani MZ, Chaudhari AS, Heckel R. Measuring Robustness in Deep Learning Based Compressive Sensing. In: Proceedings of the 38th International Conference on Machine Learning. Vol 139. Proceedings of Machine Learning Research. PMLR; 2021:2433-2444.
• Gunel B, Mardani M, Chaudhari A, Vasanawala S, and Pauly J. Weakly Supervised MR Image Reconstruction using Untrained Neural Networks. Intl Soc Magn Reson Med, (virtual), 2021

We seek to understand how convolutional neural network representations affect downstream tasks of interest, especially in scenarios where the representations can be obtained from self-supervised pre-training strategies. We explore different pre-training strategies such as inpainting and contrastive learning to determine whether such embeddings may be used for improving performance on downstream tasks such as image segmentation and determination of image quality for inverse problems such as MRI reconstruction. Furthermore, we study how building robust representations can help reduce the extent of training data for training accurate downstream models.

Select Publications:

• Dominic F, Desai A, Schmidt A, Rubin A, Gold G, Hargreaves B, and Chaudhari A. Self-Supervised Deep Learning for Knee MRI Segmentation using Limited Labeled Training Datasets. Intl Soc Magn Reson Med, (virtual), 2021.
• SSFD

As a pre-requisite to deploying models in healthcare settings, we seek to improve our understanding of why deep learning models make the predictions that they do. Specifically, we develop counterfactual explanations that seek to perturb a given test image to better interpret which model features are being activated. We explore these techniques in classification problems to curtail or exaggerate features relevant for a given class prediction, along with inverse problems such as image generation to change semantic features of images. Such counterfactuals have the potential to assist algorithm developers to better understand whether a model is utilizing spurious correlations and they can be used in a human-in-the-loop manner to assist an end-user in interpreting and acting upon on the model output.

Select Publications:

• Cohen J, Brooks R, En S, Zucker E, Pareek A, Lungren M, and Chaudhari A. Gifsplanation via Latent Shift: A Simple Autoencoder Approach to Counterfactual Generation for Chest X-rays. Medical Imaging with Deep Learning. 2021.
• Thiagarajan J, Narayanaswamy V, Rajan D, Liang J, Chaudhari A, Spanias A. Designing Counterfactual Generators using Deep Model Inversion. The Thirty-Fifth Annual Conference on Neural Information Processing Systems (NeurIPS) 2021 (virtual).