Self-supervised learning for MRI reconstruction: a review and new perspective.

The availability of huge annotated datasets and deep learning have led to astounding Natural Language Processing (NLP) progress. However, there is a plethora of low-resource languages, lack the labelled corpora to train the models in the supervision. This kind of digital divide does not allow fair NLP development across language groups. Self-supervised learning (SSL) is a new paradigm in recent years that leverages large amounts of untagged text to learn powerful representations with little or no manual annotation. This paper is a review of self-supervised deep learning models in low-resource NLP tasks. It begins with the definition of the principles of the SSL and the difference between this approach and supervised and unsupervised approaches. We describe why methods such as masked language modeling, contrastive learning, and autoregressive modeling underlie modern pre-trained transformers such as BERT, GPT, and mBERT. A particular interest is paid to multilingual and cross lingual SSL schemes which allow knowledge transfer between high-resource and low resource languages. Low-resource tasks, including machine translation, sentiment analysis, speech-to-text, and information retrieval are reviewed. As per benchmark studies, even in small labelled samples, the SSL models can achieve big gains in accuracy and generalization. Computation cost, representational bias and morphologically rich and under-documented language evaluation, are however, problematic. The researchers observe that the study suffers certain limitations, such as over-reliance on high-resource pretraining information, inequity between linguistic groups, and the difficulty of deploying large-scale SSL models in resource-constrained circumstances. The future directions include lightweight multilingual models, federation of learning in NLP and symbolic linguistic knowledge and combination with SSL. Self-supervised deep learning bridges an essential gap between the high-resource and low-resource languages, and would be a highly valuable step toward inclusive, global NLP innovation.

Electrocardiogram (ECG) is acquired during Magnetic Resonance Imaging (MRI) to monitor patients and synchronize image acquisition with the heart motion. ECG signals are highly distorted during MRI due to the complex electromagnetic environment. Automated ECG analysis is therefore complicated in this context and there is no reference technique in MRI to classify pathological heartbeats. Imaging arrhythmic patients is hence difficult in MRI. Deep Learning based heartbeat classifier have been suggested but require large databases whereas existing annotated sets of ECG in MRI are very small. We proposed a Siamese network to leverage a large database of unannotated ECGs outside MRI. This was used to develop an efficient representation of ECG signals, further used to develop a heartbeat classifier. We extensively tested several data augmentations and self-supervised learning (SSL) techniques and assessed the generalization of the obtained classifier to ECG signals acquired in MRI. These augmentations included random noises and a model simulating MRI specific artefacts. SSL pretraining improved the generalizability of heartbeat classifiers in MRI (F1=0.75) compared to Deep Learning not relying on SSL (F1=0.46) and another classical machine learning approach (F1=0.40). These promising results seem to indicate that the use of SSL techniques can learn efficient ECG signal representation, and are useful for the development of Deep Learning models even when only scarce annotated medical data are available.

This paper proposes a novel self-supervised learning framework that uses model reinforcement, REference-free LAtent map eXtraction with MOdel REinforcement (RELAX-MORE), for accelerated quantitative MRI (qMRI) reconstruction. The proposed method uses an optimization algorithm to unroll an iterative model-based qMRI reconstruction into a deep learning framework, enabling accelerated MR parameter maps that are highly accurate and robust. Unlike conventional deep learning methods which require large amounts of training data, RELAX-MORE is a subject-specific method that can be trained on single-subject data through self-supervised learning, making it accessible and practically applicable to many qMRI studies. Using quantitative mapping as an example, the proposed method was applied to the brain, knee and phantom data. The proposed method generates high-quality MR parameter maps that correct for image artifacts, removes noise, and recovers image features in regions of imperfect image conditions. Compared with other state-of-the-art conventional and deep learning methods, RELAX-MORE significantly improves efficiency, accuracy, robustness, and generalizability for rapid MR parameter mapping. This work demonstrates the feasibility of a new self-supervised learning method for rapid MR parameter mapping, that is readily adaptable to the clinical translation of qMRI.

Background Histopathological endpoints are evolving as a treatment target in Inflammatory Bowel Disease (IBD). Use of histology to screen entrants could add value in IBD clinical trials; for example, by refining eligibility criteria to ensure studies recruit patients with definitive active inflammation at the microscopic level. Several histopathological indices have been developed, but the relative complexity of available scores hinders the development of an AI algorithm without large-scale labour-intensive annotation by a pathologist. We aimed to develop computer vision tools to assist decoding the complex clinical disease features at the histological level for both Crohn’s Disease (CD) and Ulcerative Colitis (UC). This will inform understanding of disease pathology and patient stratification to support clinical trial development strategies. Methods A total of 1397 clinically annotated Haematoxylin & Eosin (H&E) images were included from 418 CD and 218 UC patients enrolled in a multicentred longitudinal Study of a Prospective Adult Research Cohort with IBD (SPARC IBD) obtained from the IBD Plexus program of the Crohn’s & Colitis Foundation. We developed an image quality control (QC) algorithm to automatically identify image and tissue processing/staining artefacts negatively impacting analysis (e.g., out-of-focus, tissue folds, overstained regions) and excluded these regions (Fig. 1). Next, a self-supervised learning (SSL) deep learning computer vison model was developed and trained to predict disease relevant features including disease diagnosis and lesional macroscopic appearance (inflammation, erosions and ulcers). Finally, to better understand the model’s predictions, we generated heatmap overlays on the tissue that show regions which the model considers to be most predictive (Fig. 2) and shared these with pathologists for qualitative evaluation. Results We find that the SSL model performs well on different downstream classification tasks such as UC vs CD (area under curve (AUC) = 0.79) and normal vs lesional tissue (AUC = 0.76). Specialist pathologist collaboration further confirmed that the heatmap overlays identified clinically relevant tissue features, including inflammatory cell infiltrates (Fig. 2). Conclusion These encouraging results support further exploration of this deep-learning algorithm to distinguish disease specific characteristics in this set of images from CD and UC patients. Further work is ongoing to validate the heatmap approach on endoscopic scores, we also plan to validate our model on IBD clinical trial datasets.

The deep learning method is an efficient solution for improving the quality of undersampled magnetic resonance (MR) image reconstruction while reducing lengthy data acquisition. Most deep learning methods neglect the mutual constraints between the real and imaginary components of complex-valued k-space data. In this paper, a new complex-valued convolutional neural network, namely, Dense-U-Dense Net (DUD-Net), is proposed to interpolate the undersampled k-space data and reconstruct MR images. The proposed network comprises dense layers, U-Net, and other dense layers in sequence. The dense layers are used to simulate the mutual constraints between real and imaginary components, and U-Net performs feature sparsity and interpolation estimation for k-space data. Two MRI datasets were used to evaluate the proposed method: brain magnitude-only MR images and knee complex-valued k-space data. Several operations were conducted for data preprocessing. First, the complex-valued MR images were synthesized by phase modulation on magnitude-only images. Second, a radial trajectory based on the golden angle was used for k-space undersampling, whereby a reversible normalization method was proposed to balance the distribution of positive and negative values in k-space data. The optimal performance of DUD-Net was demonstrated based on a quantitative evaluation of inter-method and intra-method comparisons. When compared with other methods, significant improvements were achieved, PSNRs were increased by 10.78 and 5.74dB, whereas RMSEs were decreased by 71.53% and 30.31% for magnitude and phase image, respectively. It is concluded that DUD-Net significantly improves the performance of MR image reconstruction.

In recent years, deep learning has demonstrated great potential in accelerating Magnetic Resonance Imaging (MRI). However, most existing works simply consider MRI reconstruction as an image-to-image task, or typically use a single and identical network for both frequency and image domain reconstruction simultaneously, without taking into consideration the data characteristics and internal interactions of the two domains, which leads to ineffective learning of the image reconstruction models. In this study, we proposed an attention-based, dual-domain reconstruction network to facilitate fast and accurate MRI reconstruction by adaptively learning different data characteristics in the frequency domain and image domain. In response to the inhomogeneous energy distribution in K-space, we designed a densely connected multi-attention network for frequency domain reconstruction. For image aliasing at low sampling rates, we developed a residual encoder-decoder network with an attention module for the image domain reconstruction. The overall network contains two parallel and interactive branches that simultaneously perform on K-space and image data, with the data fusion modules to fuse dual domain information in each reconstruction layer, and the long skip connections to transfer inter-layer information. We have evaluated our method on two datasets. A series of comparative and ablation experiments demonstrate that our method effectively reconstructs images from undersampled K-space data, achieving better image reconstruction performance and restoring significant textures of brain slices.

Reconstructing magnetic resonance images from undersampled k-space data is a challenging problem. This paper introduces a novel method of image reconstruction from undersampled k-space data based on the concept of singularizing operators and a novel singular k-space model. Exploring the sparsity of an image in the k-space, the singular k-space model (SKM) is proposed in terms of the k-space functions of a singularizing operator. The singularizing operator is constructed by combining basic difference operators. An algorithm is developed to reliably estimate the model parameters from undersampled k-space data. The estimated parameters are then used to recover the missing k-space data through the model, subsequently achieving high-quality reconstruction of the image using inverse Fourier transform. Experiments on physical phantom and real brain MR images have shown that the proposed SKM method constantly outperforms the popular total variation (TV) and the classical zero-filling (ZF) methods regardless of the undersampling rates, the noise levels, and the image structures. For the same objective quality of the reconstructed images, the proposed method requires much less k-space data than the TV method. The SKM method is an effective method for fast MRI reconstruction from the undersampled k-space data. Graphical abstract Two Real Images and their sparsified images by singularizing operator.

Spatiotemporal self-supervised predictive learning for atmospheric variable prediction via multi-group multi-attention

This paper introduces an efficient algorithm for magnetic resonance (MR) image reconstruction. The proposed method minimizes a linear combination of nonlocal total variation and least-square data-fitting term to reconstruct the MR images from undersampled k-space data. The nonlocal total variation is taken as the L 1-regularization functional and solved using Split Bregman iteration. The proposed algorithm is compared with previous methods in terms of the reconstruction accuracy and computational complexity. The comparison results demonstrate the superiority of the proposed algorithm for compressed MR image reconstruction.

The image reconstruction process in medical imaging can be treated as solving an inverse problem. The inverse problem is usually solved using time-consuming iterative algorithms with sparsity or other constraints. Recently, deep neural network (DNN)-based methods have been developed to accelerate the inverse-problem-solving process. However, these methods typically adopt supervised learning scheme, which requires ground truths, or labels of the solutions, for training. In many applications, it would be challenging or even impossible to obtain the ground truth, such as the tissue reflectivity function in ultrasound beamforming. In this study, a general framework based on self-supervised learning (SSL) scheme is proposed to train a DNN to solve the inverse problems. In this way, the measurements can be used as both the inputs and the labels during the training of DNN. The proposed SSL method is applied to four typical linear inverse problems for validation, i.e., plane wave ultrasound and photoacoustic image reconstructions, compressed sensing-based synthetic transmit aperture dataset recovery and deconvolution in ultrasound localization microscopy. Results show that, using the proposed framework, the trained DNN can achieve improved reconstruction accuracy with reduced computational time, compared with conventional methods.

Deep model-based magnetic resonance parameter mapping network (DOPAMINE) for fast T1 mapping using variable flip angle method.

The extended acquisition times required for MRI limit its availability in resource-constrained settings. Consequently, accelerating MRI by undersampling k-space data, which is necessary to reconstruct an image, has been a long-standing but important challenge. We aimed to develop a deep convolutional neural network (dCNN) optimisation method for MRI reconstruction and to reduce scan times and evaluate its effect on image quality and accuracy of oncological imaging biomarkers. In this multicentre, retrospective, cohort study, MRI data from patients with glioblastoma treated at Heidelberg University Hospital (775 patients and 775 examinations) and from the phase 2 CORE trial (260 patients, 1083 examinations, and 58 institutions) and the phase 3 CENTRIC trial (505 patients, 3147 examinations, and 139 institutions) were used to develop, train, and test dCNN for reconstructing MRI from highly undersampled single-coil k-space data with various acceleration rates (R=2, 4, 6, 8, 10, and 15). Independent testing was performed with MRIs from the phase 2/3 EORTC-26101 trial (528 patients with glioblastoma, 1974 examinations, and 32 institutions). The similarity between undersampled dCNN-reconstructed and original MRIs was quantified with various image quality metrics, including structural similarity index measure (SSIM) and the accuracy of undersampled dCNN-reconstructed MRI on downstream radiological assessment of imaging biomarkers in oncology (automated artificial intelligence-based quantification of tumour burden and treatment response) was performed in the EORTC-26101 test dataset. The public NYU Langone Health fastMRI brain test dataset (558 patients and 558 examinations) was used to validate the generalisability and robustness of the dCNN for reconstructing MRIs from available multi-coil (parallel imaging) k-space data. In the EORTC-26101 test dataset, the median SSIM of undersampled dCNN-reconstructed MRI ranged from 0·88 to 0·99 across different acceleration rates, with 0·92 (95% CI 0·92-0·93) for 10-times acceleration (R=10). The 10-times undersampled dCNN-reconstructed MRI yielded excellent agreement with original MRI when assessing volumes of contrast-enhancing tumour (median DICE for spatial agreement of 0·89 [95% CI 0·88 to 0·89]; median volume difference of 0·01 cm3 [95% CI 0·00 to 0·03] equalling 0·21%; p=0·0036 for equivalence) or non-enhancing tumour or oedema (median DICE of 0·94 [95% CI 0·94 to 0·95]; median volume difference of -0·79 cm3 [95% CI -0·87 to -0·72] equalling -1·77%; p=0·023 for equivalence) in the EORTC-26101 test dataset. Automated volumetric tumour response assessment in the EORTC-26101 test dataset yielded an identical median time to progression of 4·27 months (95% CI 4·14 to 4·57) when using 10-times-undersampled dCNN-reconstructed or original MRI (log-rank p=0·80) and agreement in the time to progression in 374 (95·2%) of 393 patients with data. The dCNN generalised well to the fastMRI brain dataset, with significant improvements in the median SSIM when using multi-coil compared with single-coil k-space data (p<0·0001). Deep-learning-based reconstruction of undersampled MRI allows for a substantial reduction of scan times, with a 10-times acceleration demonstrating excellent image quality while preserving the accuracy of derived imaging biomarkers for the assessment of oncological treatment response. Our developments are available as open source software and hold considerable promise for increasing the accessibility to MRI, pending further prospective validation. Deutsche Forschungsgemeinschaft (German Research Foundation) and an Else Kröner Clinician Scientist Endowed Professorship by the Else Kröner Fresenius Foundation.

As one of the main sources of remote sensing big data, synthetic aperture radar (SAR) can provide all-day and all-weather Earth image acquisition. However, speckle noise in SAR images brings a notable limitation for its big data applications, including image analysis and interpretation. Deep learning has been demonstrated as an advanced method and technology for SAR image despeckling. Most existing deep-learning-based methods adopt supervised learning and use synthetic speckled images to train the despeckling networks. This is because they need clean images as the references, and it is hard to obtain purely clean SAR images in real-world conditions. However, significant differences between synthetic speckled and real SAR images cause the domain gap problem. In other words, they cannot show superior performance for despeckling real SAR images as they do for synthetic speckled images. Inspired by recent studies on self-supervised denoising, we propose an advanced SAR image despeckling method by virtue of Bernoulli-sampling-based self-supervised deep learning, called SSD-SAR-BS. By only using real speckled SAR images, Bernoulli-sampled speckled image pairs (input–target) were obtained as the training data. Then, a multiscale despeckling network was trained on these image pairs. In addition, a dropout-based ensemble was introduced to boost the network performance. Extensive experimental results demonstrated that our proposed method outperforms the state-of-the-art for speckle noise suppression on both synthetic speckled and real SAR datasets (i.e., Sentinel-1 and TerraSAR-X).

Contrastive learning, a successful form of representational learning, has shown promising results in pretraining deep learning (DL) models for downstream tasks. When working with limited annotation data, as in medical image segmentation tasks, learning domain-specific local representations can further improve the performance of DL models. In this work, we extend the contrastive learning framework to utilize domain-specific contrast information from unlabeled Magnetic Resonance (MR) images to improve the performance of downstream MR image segmentation tasks in the presence of limited labeled data. The contrast in MR images is controlled by underlying tissue properties (e.g., T1 or T2) and image acquisition parameters. We hypothesize that learning to discriminate local representations based on underlying tissue properties should improve subsequent segmentation tasks on MR images. We propose a novel constrained contrastive learning (CCL) strategy that uses tissue-specific information via a constraint map to define positive and negative local neighborhoods for contrastive learning, embedding this information in the representational space during pretraining. For a given MR contrast image, the proposed strategy uses local signal characteristics (constraint map) across a set of related multi-contrast MR images as a surrogate for underlying tissue information. We demonstrate the utility of the approach for downstream: (1) multi-organ segmentation tasks in T2-weighted images where a DL model learns T2 information with constraint maps from a set of 2D multi-echo T2-weighted images (n=101) and (2) tumor segmentation tasks in multi-parametric images from the public brain tumor segmentation (BraTS) (n=80) dataset where DL models learn T1 and T2 information from multi-parametric BraTS images. Performance is evaluated on downstream multi-label segmentation tasks with limited data in (1) T2-weighted images of the abdomen from an in-house Radial-T2 (Train/Test=30/20), (2) public Cartesian-T2 (Train/Test=6/12) dataset, and (3) multi-parametric MR images from the public brain tumor segmentation dataset (BraTS) (Train/Test=40/50). The performance of the proposed CCL strategy is compared to state-of-the-art self-supervised contrastive learning techniques. In each task, a model is also trained using all available labeled data for supervised baseline performance. The proposed CCL strategy consistently yielded improved Dice scores, Precision, and Recall metrics, and reduced HD95 values across all segmentation tasks. We also observed performance comparable to the baseline with reduced annotation effort. The t-SNE visualization of features for T2-weighted images demonstrates its ability to embed T2 information in the representational space. On the BraTS dataset, we also observed that using an appropriate multi-contrast space to learn T1+T2, T1, or T2 information during pretraining further improved the performance of tumor segmentation tasks. Learning to embed tissue-specific information that controls MR image contrast with the proposed constrained contrastive learning improved the performance of DL models on subsequent segmentation tasks compared to conventional self-supervised contrastive learning techniques. The use of such domain-specific local representations could help understand, improve performance, and mitigate the scarcity of labeled data in MR image segmentation tasks.

Deep learning has emerged as a powerful tool in various domains, revolutionising machine learning research. However, one persistent challenge is the scarcity of labelled training data, which hampers the performance and generalisation of deep learning models. To address this limitation, researchers have developed innovative methods to overcome data scarcity and enhance deep model learning capabilities. Two prevalent techniques that have gained significant attention are transfer learning and self-supervised learning. Transfer learning leverages knowledge learned from pre-training on a large-scale dataset, such as ImageNet, and applies it to a target task with limited labelled data. This approach allows models to benefit from the learned representations and effectively transfer knowledge to new tasks, resulting in improved learning performance and generalisation. On the other hand, self-supervised learning focuses on training models using pretext tasks that do not require manual annotation, allowing them to learn valuable representations from large amounts of unlabelled data. These learned representations can then be fine-tuned for downstream tasks, mitigating the need for extensive labelled data. In recent years, transfer and self-supervised learning have found applications in various fields, including medical image processing, video recognition, and natural language processing. These approaches have demonstrated remarkable achievements, enabling breakthroughs in areas such as disease diagnosis, object recognition, and language understanding. However, while these methods offer numerous advantages, they also have limitations. For example, transfer learning may face domain mismatch issues between the pre-training and target domains, while self-supervised learning requires careful design of pretext tasks to ensure meaningful representations. This review paper explores the recent applications of these pre-training methods in various fields within the past three years. It delves into the advantages and limitations of each approach, assesses the performance of models employing these techniques, and identifies potential directions for future research. By providing a comprehensive review of current pre-training methods, this article offers guidance for selecting the best technique for specific deep learning applications to address the data scarcity issue.