Image Representation Learning Research Articles

Self-supervised learning from images: No negative pairs, no cluster-balancing

Learning with self-derived targets provides a non-contrastive method for unsupervised image representation learning, where the variety in targets is crucial. Recent work has achieved good performance by learning with targets obtained via cluster-balancing. However, the equal-cluster-size constraint becomes too restrictive for handling data with imbalanced categories or coming in small batches. In this paper, we propose a new clustering-based approach for non-contrastive image representation learning with no need for a particular architecture design or extra memory bank and no explicit constraints on cluster size. A key formulation is to learn embedding consistency and variable decorrelation in the cluster space by tweaking the batch-wise cross-correlation matrix towards an identity one. With this identitization loss incorporated, predicted cluster assignments of two randomly augmented views of the same image serve as targets for each other. We carried out comprehensive experimental studies of linear classification with learned representations of benchmark image datasets. Our results show that the proposed approach significantly outperforms state-of-the-art approaches and is more robust to class imbalance than those with cluster balancing.

Self-Supervised Representation Distribution Learning for Reliable Data Augmentation in Histopathology WSI Classification.

Multiple instance learning (MIL) based whole slide image (WSI) classification is often carried out on the representations of patches extracted from WSI with a pre-trained patch encoder. The performance of classification relies on both patch-level representation learning and MIL classifier training. Most MIL methods utilize a frozen model pre-trained on ImageNet or a model trained with self-supervised learning on histopathology image dataset to extract patch image representations and then fix these representations in the training of the MIL classifiers for efficiency consideration. However, the invariance of representations cannot meet the diversity requirement for training a robust MIL classifier, which has significantly limited the performance of the WSI classification. In this paper, we propose a Self-Supervised Representation Distribution Learning framework (SSRDL) for patch-level representation learning with an online representation sampling strategy (ORS) for both patch feature extraction and WSI-level data augmentation. The proposed method was evaluated on three datasets under three MIL frameworks. The experimental results have demonstrated that the proposed method achieves the best performance in histopathology image representation learning and data augmentation and outperforms state-of-the-art methods under different WSI classification frameworks. The code is available at https://github.com/lazytkm/SSRDL.

Rapid detection of rare events from in situX-ray diffraction data using machine learning.

High-energy X-ray diffraction methods can non-destructively map the 3D microstructure and associated attributes of metallic polycrystalline engineering materials in their bulk form. These methods are often combined with external stimuli such as thermo-mechanical loading to take snapshots of the evolving microstructure and attributes over time. However, the extreme data volumes and the high costs of traditional data acquisition and reduction approaches pose a barrier to quickly extracting actionable insights and improving the temporal resolution of these snapshots. This article presents a fully automated technique capable of rapidly detecting the onset of plasticity in high-energy X-ray microscopy data. The technique is computationally faster by at least 50 times than the traditional approaches and works for data sets that are up to nine times sparser than a full data set. This new technique leverages self-supervised image representation learning and clustering to transform massive data sets into compact, semantic-rich representations of visually salient characteristics (e.g. peak shapes). These characteristics can rapidly indicate anomalous events, such as changes in diffraction peak shapes. It is anticipated that this technique will provide just-in-time actionable information to drive smarter experiments that effectively deploy multi-modal X-ray diffraction methods spanning many decades of length scales.

Morphological profiling for drug discovery in the era of deep learning.

Morphological profiling is a valuable tool in phenotypic drug discovery. The advent of high-throughput automated imaging has enabled the capturing of a wide range of morphological features of cells or organisms in response to perturbations at the single-cell resolution. Concurrently, significant advances in machine learning and deep learning, especially in computer vision, have led to substantial improvements in analyzing large-scale high-content images at high throughput. These efforts have facilitated understanding of compound mechanism of action, drug repurposing, characterization of cell morphodynamics under perturbation, and ultimately contributing to the development of novel therapeutics. In this review, we provide a comprehensive overview of the recent advances in the field of morphological profiling. We summarize the image profiling analysis workflow, survey a broad spectrum of analysis strategies encompassing feature engineering- and deep learning-based approaches, and introduce publicly available benchmark datasets. We place a particular emphasis on the application of deep learning in this pipeline, covering cell segmentation, image representation learning, and multimodal learning. Additionally, we illuminate the application of morphological profiling in phenotypic drug discovery and highlight potential challenges and opportunities in this field.

Learning multi-granularity semantic interactive representation for joint low-light image enhancement and super-resolution

Images captured in challenging conditions often suffer from the co-existence of low contrast and low resolution. However, most joint enhancement methods focus on fitting a direct mapping from degraded images to high-quality images, which proves insufficient to handle complex degradation. To mitigate this, we propose a novel semantic prior guided interactive network (MSIRNet) to enable effective image representation learning for joint low-light enhancement and super-resolution. Specifically, a local HE-based domain transfer strategy is developed to remedy the domain gap between low-light images and the recognition scope of a generic segmentation model, thereby obtaining a rich granularity of semantic prior. To represent hybrid-scale features with semantic attributes, we propose a multi-grained semantic progressive interaction module that formulates an omnidirectional blend self-attention mechanism, facilitating deep interaction between diverse semantic knowledge and visual features. Moreover, employing our feature normalized complementary module that perceives context and cross-feature relationships, MSIRNet adaptively integrates image features with the auxiliary visual atoms provided by the Codebook, endowing the model with high-fidelity reconstruction capability. Extensive experiments demonstrate the superior performance of our MSIRNet, showing its ability to restore visually and perceptually pleasing normal-light high-resolution images.

SSiT: Saliency-Guided Self-Supervised Image Transformer for Diabetic Retinopathy Grading.

Self-supervised Learning (SSL) has been widely applied to learn image representations through exploiting unlabeled images. However, it has not been fully explored in the medical image analysis field. In this work, Saliency-guided Self-Supervised image Transformer (SSiT) is proposed for Diabetic Retinopathy (DR) grading from fundus images. We novelly introduce saliency maps into SSL, with a goal of guiding self-supervised pre-training with domain-specific prior knowledge. Specifically, two saliency-guided learning tasks are employed in SSiT: 1) Saliency-guided contrastive learning is conducted based on the momentum contrast, wherein fundus images' saliency maps are utilized to remove trivial patches from the input sequences of the momentum-updated key encoder. Thus, the key encoder is constrained to provide target representations focusing on salient regions, guiding the query encoder to capture salient features. 2) The query encoder is trained to predict the saliency segmentation, encouraging the preservation of fine-grained information in the learned representations. To assess our proposed method, four publicly-accessible fundus image datasets are adopted. One dataset is employed for pre-training, while the three others are used to evaluate the pre-trained models' performance on downstream DR grading. The proposed SSiT significantly outperforms other representative state-of-the-art SSL methods on all downstream datasets and under various evaluation settings. For example, SSiT achieves a Kappa score of 81.88% on the DDR dataset under fine-tuning evaluation, outperforming all other ViT-based SSL methods by at least 9.48%.

Hypergraph regularized nonnegative triple decomposition for multiway data analysis

Tucker decomposition is widely used for image representation, data reconstruction, and machine learning tasks, but the calculation cost for updating the Tucker core is high. Bilevel form of triple decomposition (TriD) overcomes this issue by decomposing the Tucker core into three low-dimensional third-order factor tensors and plays an important role in the dimension reduction of data representation. TriD, on the other hand, is incapable of precisely encoding similarity relationships for tensor data with a complex manifold structure. To address this shortcoming, we take advantage of hypergraph learning and propose a novel hypergraph regularized nonnegative triple decomposition for multiway data analysis that employs the hypergraph to model the complex relationships among the raw data. Furthermore, we develop a multiplicative update algorithm to solve our optimization problem and theoretically prove its convergence. Finally, we perform extensive numerical tests on six real-world datasets, and the results show that our proposed algorithm outperforms some state-of-the-art methods.

Joint low-light enhancement and deblurring with structural priors guidance

Images captured under low-light conditions usually co-exist with low light and blur degradation. Most existing cascade and joint enhancement methods may provide undesirable results, suffering from severe artifacts, deteriorating blur, and unclear details. In this paper, we propose a novel network with structural priors, including high-frequency and edge, to enable effective image representation learning for joint low-light enhancement and deblurring. Specifically, we employ a Transformer backbone to explore the global clues of the image. To compensate for the inadequate local detail optimization, we propose a multi-patch perception pyramid block that models the correlation between different size patches and ambiguity, and identifies non-uniform deblurring spatial features, facilitating the reconstruction of potential high-frequency and edge information. Furthermore, a prior-guided reconstruction block based on the parallel attention mechanism is present to adaptively correct global image with statistical features, which helps guide the model to refine sharp texture and structure. Extensive experiments performed on simulated and real-world datasets demonstrate the efficacy of our proposed method in restoring low-light blurry and low-light images with increased visual perception compared to state-of-the-art methods.

SCANeXt: Enhancing 3D medical image segmentation with dual attention network and depth-wise convolution

Existing approaches to 3D medical image segmentation can be generally categorized into convolution-based or transformer-based methods. While convolutional neural networks (CNNs) demonstrate proficiency in extracting local features, they encounter challenges in capturing global representations. In contrast, the consecutive self-attention modules present in vision transformers excel at capturing long-range dependencies and achieving an expanded receptive field. In this paper, we propose a novel approach, termed SCANeXt, for 3D medical image segmentation. Our method combines the strengths of dual attention (Spatial and Channel Attention) and ConvNeXt to enhance representation learning for 3D medical images. In particular, we propose a novel self-attention mechanism crafted to encompass spatial and channel relationships throughout the entire feature dimension. To further extract multiscale features, we introduce a depth-wise convolution block inspired by ConvNeXt after the dual attention block. Extensive evaluations on three benchmark datasets, namely Synapse, BraTS, and ACDC, demonstrate the effectiveness of our proposed method in terms of accuracy. Our SCANeXt model achieves a state-of-the-art result with a Dice Similarity Score of 95.18% on the ACDC dataset, significantly outperforming current methods.

MOINER: A Novel Multiomics Early Integration Framework for Biomedical Classification and Biomarker Discovery.

In the context of precision medicine, multiomics data integration provides a comprehensive understanding of underlying biological processes and is critical for disease diagnosis and biomarker discovery. One commonly used integration method is early integration through concatenation of multiple dimensionally reduced omics matrices due to its simplicity and ease of implementation. However, this approach is seriously limited by information loss and lack of latent feature interaction. Herein, a novel multiomics early integration framework (MOINER) based on information enhancement and image representation learning is thus presented to address the challenges. MOINER employs the self-attention mechanism to capture the intrinsic correlations of omics-features, which make it significantly outperform the existing state-of-the-art methods for multiomics data integration. Moreover, visualizing the attention embedding and identifying potential biomarkers offer interpretable insights into the prediction results. All source codes and model for MOINER are freely available https://github.com/idrblab/MOINER.

Rethinking Batch Sample Relationships for Data Representation: A Batch-Graph Transformer Based Approach

Exploring sample relationships within each mini-batch has shown great potential for learning image representations. Existing works generally adopt the regular Transformer to model the visual content relationships, ignoring the cues of semantic/label correlations between samples. Also, they generally adopt the ‘full’ self-attention mechanism which are obviously redundant and also sensitive to the noisy samples. To overcome these issues, in this paper, we design a simple yet flexible Batch-Graph Transformer (BGFormer) for mini-batch sample representations by deeply capturing the relationships of image samples from both visual and semantic perspectives. BGFormer has three main aspects. (1) It employs a flexible graph model, termed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Batch Graph</i> to jointly encode both visual and semantic relationships of samples within each mini-batch. (2) It explores the neighborhood relationships of samples by borrowing the idea of sparse graph representation which thus performs robustly, w.r.t., noisy samples. (3) It devises a novel specific Transformer architecture that mainly adopts <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dual</i> structure-constrained self-attention (SSA), together with graph normalization, FFN, etc, to carefully exploit the batch graph information for sample tokens (nodes) representations. As an application, we apply BGFormer to the metric learning tasks. Extensive experiments on four popular datasets demonstrate the effectiveness of the proposed model.

Learning image representations for anomaly detection: Application to discovery of histological alterations in drug development

We present a system for anomaly detection in histopathological images. In histology, normal samples are usually abundant, whereas anomalous (pathological) cases are scarce or not available. Under such settings, one-class classifiers trained on healthy data can detect out-of-distribution anomalous samples. Such approaches combined with pre-trained Convolutional Neural Network (CNN) representations of images were previously employed for anomaly detection (AD). However, pre-trained off-the-shelf CNN representations may not be sensitive to abnormal conditions in tissues, while natural variations of healthy tissue may result in distant representations. To adapt representations to relevant details in healthy tissue we propose training a CNN on an auxiliary task that discriminates healthy tissue of different species, organs, and staining reagents. Almost no additional labeling workload is required, since healthy samples come automatically with aforementioned labels. During training we enforce compact image representations with a center-loss term, which further improves representations for AD. The proposed system outperforms established AD methods on a published dataset of liver anomalies. Moreover, it provided comparable results to conventional methods specifically tailored for quantification of liver anomalies. We show that our approach can be used for toxicity assessment of candidate drugs at early development stages and thereby may reduce expensive late-stage drug attrition.

Depth as attention to learn image representations for visual localization, using monocular images

Image retrieval algorithms are widely used in visual localization tasks. In visual localization, we can benefit from retrieving the images depicting same landmark taken from a pose similar to the query. However, state-of-the-art image retrieval algorithms are optimized mainly for landmark retrieval, and do not take camera pose into account. To address this limitation, we propose novel Depth Attention Network (DeAttNet). DeAttNet leverages both visual and depth information in learning a global image representation. Depth varies for similar features captured from different camera poses. Based on this insight, we employ depth within an attention mechanism to discern and emphasize the salient regions. In our method, we utilize monocular depth estimation algorithms to render depth maps. Compared to RGB only image descriptors, significant improvements are obtained with the proposed method on Mapillary Street Level Sequences, Pittsburgh and Cambridge Landmark datasets.

Few-Shot Learning Meets Transformer: Unified Query-Support Transformers for Few-Shot Classification

The goal of Few-shot classification (FSL) is to identify unseen classes with very limited samples has attracted more and more attention. Usually, it is formulated as a metric learning problem. The core issue of few-shot classification is how to learn (1) consistent representations for images in both support and query sets and (2) effective metric learning for images between support and query sets. In this paper, we show that the two challenges can be well modeled simultaneously via a unified Query-Support TransFormer (QSFormer) model. To be specific, the proposed QSFormer involves <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">global</i> query-support sample Transformer (sampleFormer) branch and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">local</i> patch Transformer (patchFormer) learning branch. sampleFormer aims to capture the dependence of samples in support and query sets for image representation. It adopts the Encoder, QS-Decoder and Cross-Attention to respectively model the Support, Query (image) representation and Metric learning for few-shot classification task. Also, as a complementary to global learning branch, we adopt a local patch Transformer to extract structural representation for each image sample by capturing the long-range dependence of local image patches. In addition, we introduce a novel Cross-scale Interactive Feature Extractor (CIFE) to extract and fuse different scale CNN features as an effective backbone module for the proposed few-shot learning method. We integrate these into a unified framework and train it in an end-to-end way. A large number of experiments are conducted on four popular datasets to validate the superiority and effectiveness of the proposed QSFormer.

Attention2Minority: A salient instance inference-based multiple instance learning for classifying small lesions in whole slide images

Multiple instance learning (MIL) models have achieved remarkable success in analyzing whole slide images (WSIs) for disease classification problems. However, with regard to giga-pixel WSI classification problems, current MIL models are often incapable of differentiating a WSI with extremely small tumor lesions. This minute tumor-to-normal area ratio in a MIL bag inhibits the attention mechanism from properly weighting the areas corresponding to minor tumor lesions. To overcome this challenge, we propose salient instance inference MIL (SiiMIL), a weakly-supervised MIL model for WSI classification. We introduce a novel representation learning for histopathology images to identify representative normal keys. These keys facilitate the selection of salient instances within WSIs, forming bags with high tumor-to-normal ratios. Finally, an attention mechanism is employed for slide-level classification based on formed bags. Our results show that salient instance inference can improve the tumor-to-normal area ratio in the tumor WSIs. As a result, SiiMIL achieves 0.9225 AUC and 0.7551 recall on the Camelyon16 dataset, which outperforms the existing MIL models. In addition, SiiMIL can generate tumor-sensitive attention heatmaps that is more interpretable to pathologists than the widely used attention-based MIL method. Our experiments imply that SiiMIL can accurately identify tumor instances, which could only take up less than 1% of a WSI, so that the ratio of tumor to normal instances within a bag can increase by two to four times.

Improving Representation Learning for Histopathologic Images with Cluster Constraints.

Recent advances in whole-slide image (WSI) scanners and computational capabilities have significantly propelled the application of artificial intelligence in histopathology slide analysis. While these strides are promising, current supervised learning approaches for WSI analysis come with the challenge of exhaustively labeling high-resolution slides-a process that is both labor-intensive and timeconsuming. In contrast, self-supervised learning (SSL) pretraining strategies are emerging as a viable alternative, given that they don't rely on explicit data annotations. These SSL strategies are quickly bridging the performance disparity with their supervised counterparts. In this context, we introduce an SSL framework. This framework aims for transferable representation learning and semantically meaningful clustering by synergizing invariance loss and clustering loss in WSI analysis. Notably, our approach outperforms common SSL methods in downstream classification and clustering tasks, as evidenced by tests on the Camelyon16 and a pancreatic cancer dataset. The code and additional details are accessible at https://github.com/wwyi1828/CluSiam.

Similarity contrastive estimation for image and video soft contrastive self-supervised learning

Contrastive representation learning has proven to be an effective self-supervised learning method for images and videos. Most successful approaches are based on Noise Contrastive Estimation (NCE) and use different views of an instance as positives that should be contrasted with other instances, called negatives, that are considered as noise. However, several instances in a dataset are drawn from the same distribution and share underlying semantic information. A good data representation should contain relations between the instances, or semantic similarity and dissimilarity, that contrastive learning harms by considering all negatives as noise. To circumvent this issue, we propose a novel formulation of contrastive learning using semantic similarity between instances called Similarity Contrastive Estimation (SCE). Our training objective is a soft contrastive one that brings the positives closer and estimates a continuous distribution to push or pull negative instances based on their learned similarities. We validate empirically our approach on both image and video representation learning. We show that SCE performs competitively with the state of the art on the ImageNet linear evaluation protocol for fewer pretraining epochs and that it generalizes to several downstream image tasks. We also show that SCE reaches state-of-the-art results for pretraining video representation and that the learned representation can generalize to video downstream tasks. Source code is available here: https://github.com/juliendenize/eztorch.

Chemical structure-aware molecular image representation learning.

Current methods of molecular image-based drug discovery face two major challenges: (1) work effectively in absence of labels, and (2) capture chemical structure from implicitly encoded images. Given that chemical structures are explicitly encoded by molecular graphs (such as nitrogen, benzene rings and double bonds), we leverage self-supervised contrastive learning to transfer chemical knowledge from graphs to images. Specifically, we propose a novel Contrastive Graph-Image Pre-training (CGIP) framework for molecular representation learning, which learns explicit information in graphs and implicit information in images from large-scale unlabeled molecules via carefully designed intra- and inter-modal contrastive learning. We evaluate the performance of CGIP on multiple experimental settings (molecular property prediction, cross-modal retrieval and distribution similarity), and the results show that CGIP can achieve state-of-the-art performance on all 12 benchmark datasets and demonstrate that CGIP transfers chemical knowledge in graphs to molecular images, enabling image encoder to perceive chemical structures in images. We hope this simple and effective framework will inspire people to think about the value of image for molecular representation learning.

Unveiling hierarchical relationships for social image representation learning

Amidst the rapid propagation of images across social media platforms, the representation learning of social images has emerged as a prime research pursuit. A salient characteristic of social images lies in their multimodal properties, encompassing the visual content and textual descriptions present within individual images and the social relationships among different images. Notably, the content information within individual images and the structural information among different images exhibit a hierarchical nature. However, existing endeavors generally employ a flat framework, leading to suboptimal utilization of hierarchical relationships. Furthermore, the heterogeneity inherent in data content and structural information presents additional challenges in social image representation. In light of these challenges, we propose a novel Hierarchical Heterogeneous Graph Neural Network model for Social Image Representation learning, dubbed HHGSI. Our motivation lies in exploring and exploiting the hierarchical relationship between diverse modalities through designing the hierarchical heterogeneous network framework. HHGSI consists of an Intra-node Multimodal Graph Encoder and an Inter-node Heterogeneous Graph Neural Network to simultaneously capture fine-grained correlation within the image and heterogeneous relationships among the images. Moreover, a task-independent optimization objective is designed to make the model suitable for numerous network-oriented and multimodal tasks. Our proposal is extensively evaluated over four real-world datasets, and experimental results demonstrate the superiority of our proposal. Our code is publicly available in https://github.com/multimodal-code/HHGSI.

Deep expectation-maximization network for unsupervised image segmentation and clustering

Unsupervised learning, such as unsupervised image segmentation and clustering, are fundamental tasks in image representation learning. In this paper, we design a deep expectation-maximization (DEM) network for unsupervised image segmentation and clustering. It is based on the statistical modeling of image in its latent feature space by Gaussian mixture model (GMM), implemented in a novel deep learning framework. Specifically, in the unsupervised setting, we design an auto-encoder network and an EM module over the image latent features, for jointly learning the image latent features and GMM model of the latent features in a single framework. To construct the EM-module, we unfold the iterative operations of EM algorithm and the online EM algorithm in fixed steps to be differentiable network blocks, plugged into the network to estimate the GMM parameters of the image latent features. The proposed network parameters can be end-to-end optimized using losses based on log-likelihood of GMM, entropy of Gaussian component assignment probabilities and image reconstruction error. Extensive experiments confirm that our proposed networks achieve favorable results compared with several state-of-the-art methods in unsupervised image segmentation and clustering.