Prior work using Masked Autoencoders (MAEs) typically relies on random patch masking based on the assumption that images have significant redundancies across different channels, allowing for the reconstruction of masked content using cross-channel correlations. However, this assumption does not hold in Multi-Channel Imaging (MCI), where channels may provide complementary information with minimal feature overlap. Thus, these MAEs primarily learn local structures within individual channels from patch reconstruction, failing to fully leverage cross-channel interactions and limiting their MCI effectiveness. In this paper, we present ChA-MAEViT, an MAE-based method that enhances feature learning across MCI channels via four key strategies: (1) dynamic channel-patch masking, which compels the model to reconstruct missing channels in addition to masked patches, thereby enhancing cross-channel dependencies and improving robustness to varying channel configurations; (2) memory tokens, which serve as long-term memory aids to promote information sharing across channels, addressing the challenges of reconstructing structurally diverse channels; (3) hybrid token fusion module, which merges fine-grained patch tokens with a global class token to capture richer representations; and (4) Channel-Aware Decoder, a lightweight decoder utilizes channel tokens to effectively reconstruct image patches. Experiments on satellite and microscopy datasets, CHAMMI, JUMP-CP, and So2Sat, show that ChA-MAEViT significantly outperforms state-of-the-art MCI-ViTs by 3.0-21.5%, highlighting the importance of cross-channel interactions in MCI. Our code is publicly available at https://github.com/chaudatascience/cha_mae_vit.

Accurately quantifying cellular morphology at scale could substantially empower existing single-cell approaches. However, measuring cell morphology remains an active field of research, which has inspired multiple computer vision algorithms over the years. Here, we show that DINOv2, a vision-transformer based, self-supervised algorithm, has a remarkable ability for learning rich representations of cellular morphology without manual annotations or any other type of supervision. We apply DINOv2 to cell phenotyping problems, and compare the performance of resulting models, called Cell-DINO models, on a wide variety of tasks across two publicly available imaging datasets of diverse specifications and biological focus. Compared to supervised and other self-supervised baselines, Cell-DINO models demonstrate improved performance, especially in low annotation regimes. For instance, to classify protein localization using only 1% of annotations on a challenging single-cell dataset, Cell-DINO performs 70% better than a supervised strategy, and 24% better than another self-supervised alternative. The results show that Cell-DINO can support the study of unknown biological variation, including single-cell heterogeneity and relationships between experimental conditions, making it an excellent tool for image-based biological discovery.

Most neural networks assume that input images have a fixed number of channels (three for RGB images). However, there are many settings where the number of channels may vary, such as microscopy images where the number of channels changes depending on instruments and experimental goals. Yet, there has not been a systemic attempt to create and evaluate neural networks that are invariant to the number and type of channels. As a result, trained models remain specific to individual studies and are hardly reusable for other microscopy settings. In this paper, we present a benchmark for investigating channel-adaptive models in microscopy imaging, which consists of 1) a dataset of varied-channel single-cell images, and 2) a biologically relevant evaluation framework. In addition, we adapted several existing techniques to create channel-adaptive models and compared their performance on this benchmark to fixed-channel, baseline models. We find that channel-adaptive models can generalize better to out-of-domain tasks and can be computationally efficient. We contribute a curated dataset and an evaluation API to facilitate objective comparisons in future research and applications.

Data is one of the most important keys for success when studying a simple or a complex phenomenon. With the use of deep learning exploding and its democratization, non-computer science experts may struggle to use highly complex deep learning architectures, even when straightforward models offer them suitable performances. In this article, we study the specific and common problem of data imbalance in real databases as most of the bad performance problems are due to the data itself. We cover two keys aspects. First, we propose ways to deal with the situation when a data set contains different levels of imbalance. Classical imbalanced learning strategies cannot be directly applied when using multi-attribute deep learning models, i.e., multi-task or multi-label architectures. Therefore, one of our contributions is a proposed adaptation to face each one of the problems derived from imbalance. Second, we demonstrate that with little to no imbalance, straightforward deep learning models work well. However, for non-experts, these models can be seen as black boxes, where all efforts are invested in preprocessing the data. To simplify the problem, we perform the classification task without features that are costly to extract, such as part localization which is widely used in the state of the art of attribute classification. We make use of three widely known attribute databases, CUB-200-2011 - CUB as our main use case due to its deeply imbalanced nature, along with two balanced databases: celebA and AwA2. All of them contain multi-attribute annotations. The results of very fine-grained attribute learning demonstrate that in the presence of imbalance, our proposed strategies make it possible to have competitive results against the state of the art while taking advantage of multi-attribute deep learning models. We also noticed an increase in performance while using a specialized loss function (Focal Loss). For CUB, we have competitive results, and for CelebA and AwA2 our strategies over-perform the state of the art.

We study the problem of learning representations for single cells in microscopy images to discover biological relationships between their experimental conditions. Many new applications in drug discovery and functional genomics require capturing the morphology of individual cells as comprehensively as possible. Deep convolutional neural networks (CNNs) can learn powerful visual representations, but require ground truth for training; this is rarely available in biomedical profiling experiments. While we do not know which experimental treatments produce cells that look alike, we do know that cells exposed to the same experimental treatment should generally look similar. Thus, we explore training CNNs using a weakly supervised approach that uses this information for feature learning. In addition, the training stage is regularized to control for unwanted variations using mixup or RNNs. We conduct experiments on two different datasets; the proposed approach yields single-cell embeddings that are more accurate than the widely adopted classical features, and are competitive with previously proposed transfer learning approaches.

We present an active detection model for localizing objects in scenes. The model is class-specific and allows an agent to focus attention on candidate regions for identifying the correct location of a target object. This agent learns to deform a bounding box using simple transformation actions, with the goal of determining the most specific location of target objects following top-down reasoning. The proposed localization agent is trained using deep reinforcement learning, and evaluated on the Pascal VOC 2007 dataset. We show that agents guided by the proposed model are able to localize a single instance of an object after analyzing only between 11 and 25 regions in an image, and obtain the best detection results among systems that do not use object proposals for object localization.

The Flickr30k dataset has become a standard benchmark for sentence-based image description. This paper presents Flickr30k Entities, which augments the 158k captions from Flickr30k with 244k coreference chains linking mentions of the same entities in images, as well as 276k manually annotated bounding boxes corresponding to each entity. Such annotation is essential for continued progress in automatic image description and grounded language understanding. We present experiments demonstrating the usefulness of our annotations for text-to-image reference resolution, or the task of localizing textual entity mentions in an image, and for bidirectional image-sentence retrieval. These experiments confirm that we can further improve the accuracy of state-of-the-art retrieval methods by training with explicit region-to-phrase correspondence, but at the same time, they show that accurately inferring this correspondence given an image and a caption remains really challenging.

This paper presents BIGS the Big Image Data Analysis Toolkit, a software framework for large scale image processing and analysis over heterogeneous computing resources, such as those available in clouds, grids, computer clusters or throughout scattered computer resources (desktops, labs) in an opportunistic manner. Through BIGS, eScience for image processing and analysis is conceived to exploit coarse grained parallelism based on data partitioning and parameter sweeps, avoiding the need of inter-process communication and, therefore, enabling loosely coupled computing nodes (BIGS workers). It adopts an uncommitted resource allocation model where (1) experimenters define their image processing pipelines in a simple configuration file, (2) a schedule of jobs is generated and (3) workers, as they become available, take over pending jobs as long as their dependency on other jobs is fulfilled. BIGS workers act autonomously, querying the job schedule to determine which one to take over. This removes the need for a central scheduling node, requiring only access by all workers to a shared information source. Furthermore, BIGS workers are encapsulated within different technologies to enable their agile deployment over the available computing resources. Currently they can be launched through the Amazon EC2 service over their cloud resources, through Java Web Start from any desktop computer and through regular scripting or SSH commands. This suits well different kinds of research environments, both when accessing dedicated computing clusters or clouds with committed computing capacity or when using opportunistic computing resources whose access is seldom or cannot be provisioned in advance. We also adopt a NoSQL storage model to ensure the scalability of the shared information sources required by all workers, including within BIGS support for HBase and Amazon's DynamoDB service. Overall, BIGS now enables researchers to run large scale image processing pipelines in an easy, affordable and unplanned manner with the capability to take over computing resources as they become available at run time. This is shown in this paper by using BIGS in different experimental setups in the Amazon cloud and in an opportunistic manner, demonstrating its configurability, adaptability and scalability capabilities.

Image collections on the internet and other sources of information can naturally include attached text descriptions. This work considers the problem of fusing two data modalities: visual content and text keywords, to allow a flexible image indexing scheme. The proposed strategy learns multimodal relationships using matrix reconstruction principles and factorization algorithms, allowing one data modality to be represented in another modality space. We further exploit this exchangeability property, to fuse the modalities in any of the representation spaces by backprojecting predicted data to the input space. An experimental evaluation was carried out on the Corel 5K and MIRFlickr data sets using example images without text as query paradigm. Experimental results demonstrate the ability of the proposed strategy to find multimodal links between data and make them useful to improve the image retrieval performance.

While most existing enhancement tools for photographs have universal auto-enhancement functionality, recent research shows that users can have personalized preferences. In this paper, we explore whether such personalized preferences in image enhancement tend to cluster and whether users can be grouped according to such preferences. To this end, we analyze a comprehensive data set of image enhancements collected from 336 users via Amazon Mechanical Turk. We find that such clusters do exist and can be used to derive methods to learn statistical preference models from a group of users. We also present a probabilistic framework that exploits the ideas behind collaborative filtering to automatically enhance novel images for new users. Experiments show that inferring clusters in image enhancement preferences results in better prediction of image enhancement preferences and outperforms generic auto-correction tools.

Visual content has become an important component of the web. In many cases, visual content is mixed with other modalities (e.g. text) that can be exploited to extract information and knowledge. This paper presents a strategy for mining multimodal visual content. The strategy encompasses two main components: a rich representation of the multimodal objects and a model for automatically annotating unannotated images. The proposed method has two distinguishing characteristics: it uses a bag-of-features representation for images and a non-negative matrix factorization algorithm to build a latent representation.