Self-supervised Method Research Articles

A self-supervised learning network for remote heart rate measurement

The rPPG technology is a facial video-based non-contact physiological parameter detection technology, which has been widely used in the measurement of heart rate, respiration, blood oxygen and other physiological parameters. Most of the current research hotspots in this field are based on deep learning methods, which typically necessitate a significant amount of data samples for training the neural network to obtain accurate measurement results. However, the high cost required to label the data samples limits the diffusion and application of the technique. To address the above problems, we propose a new self-supervised aimed at acquiring the ability to estimate rPPG signals from facial videos, eliminating the need for labeled data. The framework expands the unlabeled video samples into multiple positive/negative samples and uses contrast learning to obtain an rPPG signal estimation network, which outputs the corresponding rPPG signals of face videos. To promote the convergence of network training, a new frequency loss function is designed. This function can effectively shorten the distance between the frequencies of similar sample signals and push the distance between the frequencies of different sample signals far away, so as to enhance the frequency consistency between sample signals and make the model easier to learn the differences between different sample pairs. Our method is evaluated on four standard rPPG datasets, the experimental results show that the accuracy is close to the current best supervised method, and is superior to the previous self-supervised method without using any labeled data.

Self-Supervised Learning for Improved Optical Coherence Tomography Detection of Macular Telangiectasia Type 2

Deep learning image analysis often depends on large, labeled datasets, which are difficult to obtain for rare diseases. To develop a self-supervised approach for automated classification of macular telangiectasia type 2 (MacTel) on optical coherence tomography (OCT) with limited labeled data. This was a retrospective comparative study. OCT images from May 2014 to May 2019 were collected by the Lowy Medical Research Institute, La Jolla, California, and the University of Washington, Seattle, from January 2016 to October 2022. Clinical diagnoses of patients with and without MacTel were confirmed by retina specialists. Data were analyzed from January to September 2023. Two convolutional neural networks were pretrained using the Bootstrap Your Own Latent algorithm on unlabeled training data and fine-tuned with labeled training data to predict MacTel (self-supervised method). ResNet18 and ResNet50 models were also trained using all labeled data (supervised method). The ground truth yes vs no MacTel diagnosis is determined by retinal specialists based on spectral-domain OCT. The models' predictions were compared against human graders using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under precision recall curve (AUPRC), and area under the receiver operating characteristic curve (AUROC). Uniform manifold approximation and projection was performed for dimension reduction and GradCAM visualizations for supervised and self-supervised methods. A total of 2636 OCT scans from 780 patients with MacTel and 131 patients without MacTel were included from the MacTel Project (mean [SD] age, 60.8 [11.7] years; 63.8% female), and another 2564 from 1769 patients without MacTel from the University of Washington (mean [SD] age, 61.2 [18.1] years; 53.4% female). The self-supervised approach fine-tuned on 100% of the labeled training data with ResNet50 as the feature extractor performed the best, achieving an AUPRC of 0.971 (95% CI, 0.969-0.972), an AUROC of 0.970 (95% CI, 0.970-0.973), accuracy of 0.898%, sensitivity of 0.898, specificity of 0.949, PPV of 0.935, and NPV of 0.919. With only 419 OCT volumes (185 MacTel patients in 10% of labeled training dataset), the ResNet18 self-supervised model achieved comparable performance, with an AUPRC of 0.958 (95% CI, 0.957-0.960), an AUROC of 0.966 (95% CI, 0.964-0.967), and accuracy, sensitivity, specificity, PPV, and NPV of 90.2%, 0.884, 0.916, 0.896, and 0.906, respectively. The self-supervised models showed better agreement with the more experienced human expert graders. The findings suggest that self-supervised learning may improve the accuracy of automated MacTel vs non-MacTel binary classification on OCT with limited labeled training data, and these approaches may be applicable to other rare diseases, although further research is warranted.

Self-supervised context-aware correlation filter for robust landmark tracking in liver ultrasound sequences.

Respiratory motion-induced displacement of internal organs poses a significant challenge in image-guided radiation therapy, particularly affecting liver landmark tracking accuracy. Addressing this concern, we propose a self-supervised method for robust landmark tracking in long liver ultrasound sequences. Our approach leverages a Siamese-based context-aware correlation filter network, trained by using the consistency loss between forward tracking and back verification. By effectively utilizing both labeled and unlabeled liver ultrasound images, our model, Siam-CCF , mitigates the impact of speckle noise and artifacts on ultrasonic image tracking by a context-aware correlation filter. Additionally, a fusion strategy for template patch feature helps the tracker to obtain rich appearance information around the point-landmark. Siam-CCF achieves a mean tracking error of 0.79±0.83 mm at a frame rate of 118.6 fps, exhibiting a superior speed-accuracy trade-off on the public MICCAI 2015 Challenge on Liver Ultrasound Tracking (CLUST2015) 2D dataset. This performance won the 5th place on the CLUST2015 2D point-landmark tracking task. Extensive experiments validate the effectiveness of our proposed approach, establishing it as one of the top-performing techniques on the CLUST2015 online leaderboard at the time of this submission.

Thyroid ultrasound diagnosis improvement via multi-view self-supervised learning and two-stage pre-training

Thyroid nodule classification and segmentation in ultrasound images are crucial for computer-aided diagnosis; however, they face limitations owing to insufficient labeled data. In this study, we proposed a multi-view contrastive self-supervised method to improve thyroid nodule classification and segmentation performance with limited manual labels. Our method aligns the transverse and longitudinal views of the same nodule, thereby enabling the model to focus more on the nodule area. We designed an adaptive loss function that eliminates the limitations of the paired data. Additionally, we adopted a two-stage pre-training to exploit the pre-training on ImageNet and thyroid ultrasound images. Extensive experiments were conducted on a large-scale dataset collected from multiple centers. The results showed that the proposed method significantly improves nodule classification and segmentation performance with limited manual labels and outperforms state-of-the-art self-supervised methods. The two-stage pre-training also significantly exceeded ImageNet pre-training.

Choice of training label matters: how to best use deep learning for quantitative MRI parameter estimation

Deep learning (DL) is gaining popularity as a parameter estimation method for quantitative MRI. A range of competing implementations have been proposed, relying on either supervised or self-supervised learning. Self-supervised approaches, sometimes referred to as unsupervised, have been loosely based on auto-encoders, whereas supervised methods have, to date, been trained on groundtruth labels. These two learning paradigms have been shown to have distinct strengths. Notably, self-supervised approaches offer lower-bias parameter estimates than their supervised alternatives. This result is counterintuitive – incorporating prior knowledge with supervised labels should, in theory, lead to improved accuracy. In this work, we show that this apparent limitation of supervised approaches stems from the naïve choice of groundtruth training labels. By training on labels which are deliberately not groundtruth, we show that the low-bias parameter estimation previously associated with self-supervised methods can be replicated – and improved on – within a supervised learning framework. This approach sets the stage for a single, unifying, deep learning parameter estimation framework, based on supervised learning, where trade-offs between bias and variance are made by careful adjustment of training label.

Knowledge-integrated Multi-modal Movie Turning Point Identification

The rapid development of artificial intelligence provides rich technologies and tools for the automated understanding of literary works. As a comprehensive carrier of storylines, movies are natural multimodal data sources that provide sufficient data foundations, and how to fully leverage the benefits of data remains a sustainable research hotspot. In addition, the efficient representation of multi-source data also poses new challenges for information fusion technology. Therefore, we propose a knowledge-enhanced turning points identification (KTPi) method for multimodal scene recognition. First, the BiLSTM method is used to encode scene text and integrate contextual information into scene representations to complete text sequence modeling. Then, the graph structure is used to model all scenes, which strengthens long-range semantic dependencies between scenes and enhances scene representations using graph convolution network. After, the self-supervised method is used to obtain the optimal number of neighboring nodes in sparse graph. Next, actor and verb knowledge involved in the scene text are added to the multimodal data to enhance the diversity of scene feature expressions. Finally, the teacher-student network strategy is used to train the KTPi model. Experimental results show that KTPi outperforms baseline methods in scene role recognition tasks, and ablation experiments show that incorporating knowledge into multimodal model can improve its performance.

Fully Self-Supervised Out-of-Domain Few-Shot Learning with Masked Autoencoders.

Few-shot learning aims to identify unseen classes with limited labelled data. Recent few-shot learning techniques have shown success in generalizing to unseen classes; however, the performance of these techniques has also been shown to degrade when tested on an out-of-domain setting. Previous work, additionally, has also demonstrated increasing reliance on supervised finetuning in an off-line or online capacity. This paper proposes a novel, fully self-supervised few-shot learning technique (FSS) that utilizes a vision transformer and masked autoencoder. The proposed technique can generalize to out-of-domain classes by finetuning the model in a fully self-supervised method for each episode. We evaluate the proposed technique using three datasets (all out-of-domain). As such, our results show that FSS has an accuracy gain of 1.05%, 0.12%, and 1.28% on the ISIC, EuroSat, and BCCD datasets, respectively, without the use of supervised training.

Debiased Visual Question Answering via the perspective of question types

Visual Question Answering (VQA) aims to answer questions according to the given image. However, current VQA models tend to rely solely on textual information from the questions and ignore the visual information in the images to get answers, which is caused by bias that is generated during the training phase. Previous studies have shown that bias in VQA is mainly caused by the text modality, and our analysis suggests that question type is a crucial factor in bias formation. To address this bias, we proposed a self-supervised method including the Against Biased Samples(ABS) module that performs targeted debiasing by selecting samples that are prone to bias, and the Shuffle Question types(SQT) module that constructs negative samples by randomly replacing the question types of the samples selected by the ABS, to interrupting the shortcuts from question type to answer. Our approach mitigates the question-to-answer bias without using external annotations, overcoming the prior language problem. Additionally, we designed a new objective function for negative samples. Experimental results indicate that our method outperforms both self-supervised-based and supervised-based state-of-the-art approaches, achieving 70.36% accuracy on the VQA-CP v2 dataset.

Network depth affects inference of gene sets from bacterial transcriptomes using denoising autoencoders.

The increasing number of publicly available bacterial gene expression data sets provides an unprecedented resource for the study of gene regulation in diverse conditions, but emphasizes the need for self-supervised methods for the automated generation of new hypotheses. One approach for inferring coordinated regulation from bacterial expression data is through neural networks known as denoising autoencoders (DAEs) which encode large datasets in a reduced bottleneck layer. We have generalized this application of DAEs to include deep networks and explore the effects of network architecture on gene set inference using deep learning. We developed a DAE-based pipeline to extract gene sets from transcriptomic data in Escherichia coli, validate our method by comparing inferred gene sets with known pathways, and have used this pipeline to explore how the choice of network architecture impacts gene set recovery. We find that increasing network depth leads the DAEs to explain gene expression in terms of fewer, more concisely defined gene sets, and that adjusting the width results in a tradeoff between generalizability and biological inference. Finally, leveraging our understanding of the impact of DAE architecture, we apply our pipeline to an independent uropathogenic E.coli dataset to identify genes uniquely induced during human colonization. https://github.com/BarquistLab/DAE_architecture_exploration.

SENSE: Self-Evolving Learning for Self-Supervised Monocular Depth Estimation.

Self-supervised depth estimation methods can achieve competitive performance using only unlabeled monocular videos, but they suffer from the uncertainty of jointly learning depth and pose without any ground truths of both tasks. Supervised framework provides robust and superior performance but is limited by the scope of the labeled data. In this paper, we introduce SENSE, a novel learning paradigm for self-supervised monocular depth estimation that progressively evolves the prediction result using supervised learning, but without requiring labeled data. The key contribution of our approach stems from the novel use of the pseudo labels - the noisy depth estimation from the self-supervised methods. We surprisingly find that a fully supervised depth estimation network trained using the pseudo labels can produce even better results than its "ground truth". To push the envelope further, we then evolve the self-supervised backbone by replacing its depth estimation branch with that fully supervised network. Based on this idea, we devise a comprehensive training pipeline that alternatively enhances the two key branches (depth and pose estimation) of the self-supervised backbone network. Our proposed approach can effectively ease the difficulty of multi-task training in self-supervised depth estimation. Experimental results have shown that our proposed approach achieves state-of-the-art results on the KITTI dataset.

TriCI: Triple Cross-Intra Branch Contrastive Learning for Point Cloud Analysis.

Whereas contrastive learning eliminates the need for labeled data, existing methods may suffer from inadequate features due to the conventional single shared encoder structure and struggle to fully harness the rich spectrum of 3D augmentations. In this paper, we propose TriCI, a self-supervised method that designs a triple-branch contrastive learning architecture. During contrastive pre-training, we generate three augmented versions of each input point cloud sample and pair each augmented sample with the original one, resulting in three unique positive pairs. We subsequently feed the pairs into three distinct encoders, each of which extracts features from its corresponding input positive pair. We design a novel cross-branch contrastive loss and use it along with the intra-branch contrastive loss to jointly train our network. The proposed cross-branch loss effectively aligns the output features from different perspectives for pre-training and facilitates their integration for downstream tasks, particularly in object-level scenarios. The intra-branch loss helps maximize the feature correspondences within positive pairs. Extensive experiments demonstrate the superiority of our TriCI in self-supervised learning, and show its strong ability in enhancing the performance of downstream object classification and part segmentation tasks. Interestingly, our TriCI achieves a 92.9% accuracy for linear SVM evaluation on ModelNet40, exceeding its closest competitor by 1.7% and even exceeding some supervised methods.

Self-Supervised 3D Scene Flow Estimation and Motion Prediction using Local Rigidity Prior.

In this article, we investigate self-supervised 3D scene flow estimation and class-agnostic motion prediction on point clouds. A realistic scene can be well modeled as a collection of rigidly moving parts, therefore its scene flow can be represented as a combination of rigid motion of these individual parts. Building upon this observation, we propose to generate pseudo scene flow labels for self-supervised learning through piecewise rigid motion estimation, in which the source point cloud is decomposed into local regions and each region is treated as rigid. By rigidly aligning each region with its potential counterpart in the target point cloud, we obtain a region-specific rigid transformation to generate its pseudo flow labels. To mitigate the impact of potential outliers on label generation, when solving the rigid registration for each region, we alternately perform three steps: establishing point correspondences, measuring the confidence for the correspondences, and updating the rigid transformation based on the correspondences and their confidence. As a result, confident correspondences will dominate label generation, and a validity mask will be derived for the generated pseudo labels. By using the pseudo labels together with their validity mask for supervision, models can be trained in a self-supervised manner. Extensive experiments on FlyingThings3D and KITTI datasets demonstrate that our method achieves new state-of-the-art performance in self-supervised scene flow learning, without any ground truth scene flow for supervision, even performing better than some supervised counterparts. Additionally, our method is further extended to class-agnostic motion prediction and significantly outperforms previous state-of-the-art self-supervised methods on nuScenes dataset.

DSMT-Net: Dual Self-Supervised Multi-Operator Transformation for Multi-Source Endoscopic Ultrasound Diagnosis.

Pancreatic cancer has the worst prognosis of all cancers. The clinical application of endoscopic ultrasound (EUS) for the assessment of pancreatic cancer risk and of deep learning for the classification of EUS images have been hindered by inter-grader variability and labeling capability. One of the key reasons for these difficulties is that EUS images are obtained from multiple sources with varying resolutions, effective regions, and interference signals, making the distribution of the data highly variable and negatively impacting the performance of deep learning models. Additionally, manual labeling of images is time-consuming and requires significant effort, leading to the desire to effectively utilize a large amount of unlabeled data for network training. To address these challenges, this study proposes the Dual Self-supervised Multi-Operator Transformation Network (DSMT-Net) for multi-source EUS diagnosis. The DSMT-Net includes a multi-operator transformation approach to standardize the extraction of regions of interest in EUS images and eliminate irrelevant pixels. Furthermore, a transformer-based dual self-supervised network is designed to integrate unlabeled EUS images for pre-training the representation model, which can be transferred to supervised tasks such as classification, detection, and segmentation. A large-scale EUS-based pancreas image dataset (LEPset) has been collected, including 3,500 pathologically proven labeled EUS images (from pancreatic and non-pancreatic cancers) and 8,000 unlabeled EUS images for model development. The self-supervised method has also been applied to breast cancer diagnosis and was compared to state-of-the-art deep learning models on both datasets. The results demonstrate that the DSMT-Net significantly improves the accuracy of pancreatic and breast cancer diagnosis.

Efficient 3D Representation Learning for Medical Image Analysis

Machine learning approaches have significantly advanced the 3D medical images analysis, such as the CT and MRI scans, which enables improved diagnosis and treatment evaluation. These image volumes provide rich spatial context for understanding the internal brain and body anatomies. Typical medical image analysis tasks, such as segmentation, reconstruction and registration, are essential for characterizing this context. Related to 3D data formats, meshes, point clouds and others are used to represent the anatomical structures, each with unique applications. To better capture the spatial information and address data scarcity, self- and semi-supervised learning methods have emerged. However, efficient 3D representation learning remains challenging. Recently, Transformers have shown promise, leveraging the self-attention mechanisms that perform well on transfer learning and self-supervised methods. These techniques are applied for medical domains without extensive manual labeling. This work explores data-efficient models, scalable deep learning, semantic context utilization and transferability in 3D medical image analysis. We also evaluated the foundational models, self-supervised pre- training, transfer learning and prompt tuning, thus advancing this critical field.

Self-Supervised Video Representation Learning by Video Incoherence Detection.

This article introduces a novel self-supervised method that leverages incoherence detection for video representation learning. It stems from the observation that the visual system of human beings can easily identify video incoherence based on their comprehensive understanding of videos. Specifically, we construct the incoherent clip by multiple subclips hierarchically sampled from the same raw video with various lengths of incoherence. The network is trained to learn the high-level representation by predicting the location and length of incoherence given the incoherent clip as input. Additionally, we introduce intravideo contrastive learning to maximize the mutual information between incoherent clips from the same raw video. We evaluate our proposed method through extensive experiments on action recognition and video retrieval using various backbone networks. Experiments show that our proposed method achieves remarkable performance across different backbone networks and different datasets compared to previous coherence-based methods.

Self-Supervised Temporal Graph Learning With Temporal and Structural Intensity Alignment.

Temporal graph learning aims to generate high-quality representations for graph-based tasks with dynamic information, which has recently garnered increasing attention. In contrast to static graphs, temporal graphs are typically organized as node interaction sequences over continuous time rather than an adjacency matrix. Most temporal graph learning methods model current interactions by incorporating historical neighborhood. However, such methods only consider first-order temporal information while disregarding crucial high-order structural information, resulting in suboptimal performance. To address this issue, we propose a self-supervised method called S2T for temporal graph learning, which extracts both temporal and structural information to learn more informative node representations. Notably, the initial node representations combine first-order temporal and high-order structural information differently to calculate two conditional intensities. An alignment loss is then introduced to optimize the node representations, narrowing the gap between the two intensities and making them more informative. Concretely, in addition to modeling temporal information using historical neighbor sequences, we further consider structural knowledge at both local and global levels. At the local level, we generate structural intensity by aggregating features from high-order neighbor sequences. At the global level, a global representation is generated based on all nodes to adjust the structural intensity according to the active statuses on different nodes. Extensive experiments demonstrate that the proposed model S2T achieves at most 10.13% performance improvement compared with the state-of-the-art competitors on several datasets.

Self-Supervised Medical Image Denoising Based on WISTA-Net for Human Healthcare in Metaverse.

Medical image processing plays an important role in the interaction of real world and metaverse for healthcare. Self-supervised denoising based on sparse coding methods, without any prerequisite on large-scale training samples, has been attracting extensive attention for medical image processing. Whereas, existing self-supervised methods suffer from poor performance and low efficiency. In this paper, to achieve state-of-the-art denoising performance on the one hand, we present a self-supervised sparse coding method, named the weighted iterative shrinkage thresholding algorithm (WISTA). It does not rely on noisy-clean ground-truth image pairs to learn from only a single noisy image. On the other hand, to further improve denoising efficiency, we unfold the WISTA to construct a deep neural network (DNN) structured WISTA, named WISTA-Net. Specifically, in WISTA, motivated by the merit of the lp-norm, WISTA-Net has better denoising performance than the classical orthogonal matching pursuit (OMP) algorithm and the ISTA. Moreover, leveraging the high-efficiency of DNN structure in parameter updating, WISTA-Net outperforms the compared methods in denoising efficiency. In detail, for a 256 by 256 noisy image, the running time of WISTA-Net is 4.72 s on the CPU, which is much faster than WISTA, OMP, and ISTA by 32.88 s, 13.06 s, and 6.17 s, respectively.

Compact Goal Representation Learning via Information Bottleneck in Goal-Conditioned Reinforcement Learning.

We propose an Information bottleneck (IB) for Goal representation learning (InfoGoal), a self-supervised method for generalizable goal-conditioned reinforcement learning (RL). Goal-conditioned RL learns a policy from reward signals to predict actions for reaching desired goals. However, the policy would overfit the task-irrelevant information contained in the goal and may be falsely or ineffectively generalized to reach other goals. A goal representation containing sufficient task-relevant information and minimum task-irrelevant information is guaranteed to reduce generalization errors. However, in goal-conditioned RL, it is difficult to balance the tradeoff between task-relevant information and task-irrelevant information because of the sparse and delayed learning signals, i.e., reward signals, and the inevitable task-relevant information sacrifice caused by information compression. Our InfoGoal learns a minimum and sufficient goal representation with dense and immediate self-supervised learning signals. Meanwhile, InfoGoal adaptively adjusts the weight of information minimization to achieve maximum information compression with a reasonable sacrifice of task-relevant information. Consequently, InfoGoal enables policy to generate a targeted trajectory toward states where the desired goal can be found with high probability and broadly explores those states. We conduct experiments on both simulated and real-world tasks, and our method significantly outperforms baseline methods in terms of policy optimality and the success rate of reaching unseen test goals. Video demos are available at infogoal.github.io.

Generalizable stereo depth estimation with masked image modelling.

Generalizable and accurate stereo depth estimation is vital for 3D reconstruction, especially in surgery. Supervised learning methods obtain best performance however, limited ground truth data for surgical scenes limits generalizability. Self-supervised methods don't need ground truth, but suffer from scale ambiguity and incorrect disparity prediction due to inconsistency of photometric loss. This work proposes a two-phase training procedure that is generalizable and retains the high performance of supervised methods. It entails: (1) performing self-supervised representation learning of left and right views via masked image modelling (MIM) to learn generalizable semantic stereo features (2) utilizing the MIM pre-trained model to learn robust depth representation via supervised learning for disparity estimation on synthetic data only. To improve stereo representations learnt via MIM, perceptual loss terms are introduced, which improve the model's stereo representations learnt by explicitly encouraging the learning of higher scene-level features. Qualitative and quantitative performance evaluation on surgical and natural scenes shows that the approach achieves sub-millimetre accuracy and lowest errors respectively, setting a new state-of-the-art. Despite not training on surgical nor natural scene data for disparityestimation.

Cross-View Recurrence-Based Self-Supervised Super-Resolution of Light Field

Compared with external-supervised learning-based (ESLB) methods, self-supervised learning-based (SSLB) methods can overcome the domain gap problem caused by different light field (LF) acquisition conditions, which results in the performance degradation of light field super-resolution on unseen test datasets. Current SSLB methods exploit the cross-scale recurrence feature in the single view image for super-resolution, ignoring the correlation information among views. Different from previous works, we propose a cross-view recurrence-based self-supervised mapping framework to correlate complementary information among views in the down-scaled input LF. Specifically, the cross-view recurrence information consists of geometry structure features and similar structure features. The former is to provide sub-pixel information according to disparity correlations among adjacent views, and the latter is to acquire similar color and contour information among arbitrary views, which can compensate for error disparity guidance of geometry structure features in sharp variance areas. Moreover, instead of the widely used “All-to-All” strategy, we propose a “Part-to-Part” mapping strategy, which is better competent for SSLB approaches with limited training examples solely extracted from input LF. Finally, considering that self-supervised methods need to retrain from the beginning toward each test image, based on the proposed “part-to-part” strategy, an efficient end-to-end network is designed to extract these cross-view features for superior SASR performance with less training time. Experiment results demonstrate that our method outperforms other state-of-the-art ESLB methods on both large and small domain gap cases. Compared with the only SSLB method (LFZSSR [1]), our approach achieves better performance with 524 times less training time.