Scholarly Publications - UMRAM

Permanent URI for this collectionhttps://hdl.handle.net/11693/115674

Browse

Browsing Scholarly Publications - UMRAM by Author "Acar, Mert"

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    Segmentation-aware MRI reconstruction
    (Springer Cham, 2022-09-22) Acar, Mert; Çukur, Tolga; Öksüz, İ.
    Deep learning models have been broadly adopted for accelerating MRI acquisitions in recent years. A common approach is to train deep models based on loss functions that place equal emphasis on reconstruction errors across the field-of-view. This homogeneous weighting of loss contributions might be undesirable in cases where the diagnostic focus is on tissues in a specific subregion of the image. In this paper, we propose a framework for segmentation-aware reconstruction based on segmentation as a proxy task. We leverage an end-to-end model comprising reconstruction and segmentation networks; and leverage backpropagation of segmentation error to devise a pseudo-attention effect to focus the reconstruction network. We introduce a novel stabilization method to prevent convergence onto a local minima with unacceptably poor reconstruction or segmentation performance. Our stabilization approach initiates learning on fully-sampled acquisitions, and gradually increases the undersampling rate assumed in the training set to its desired value. We validate our approach for cardiac MR reconstruction on the publicly available OCMR dataset. Segmentation-aware reconstruction significantly outperforms vanilla reconstruction for cardiac imaging.
  • Thumbnail Image
    ItemOpen Access
    Self-supervised dynamic MRI reconstruction
    (Springer, 2021-09-25) Acar, Mert; Çukur, Tolga; Öksüz, İlkay
    Deep learning techniques have recently been adopted for accelerating dynamic MRI acquisitions. Yet, common frameworks for model training rely on availability of large sets of fully-sampled MRI data to construct a ground-truth for the network output. This heavy reliance is undesirable as it is challenging to collect such large datasets in many applications, and even impossible for high spatiotemporal-resolution protocols. In this paper, we introduce self-supervised training to deep neural architectures for dynamic reconstruction of cardiac MRI. We hypothesize that, in the absence of ground-truth data, elevating complexity in self-supervised models can instead constrain model performance due to the deficiencies in training data. To test this working hypothesis, we adopt self-supervised learning on recent state-of-the-art deep models for dynamic MRI, with varying degrees of model complexity. Comparison of supervised and self-supervised variants of deep reconstruction models reveals that compact models have a remarkable advantage in reliability against performance loss in self-supervised settings.