BACKGROUND: The use of machine learning in medical diagnosis and treatment has grown significantly in recent years with the development of computer-aided diagnosis systems, often based on annotated medical radiology images. However, the lack of large annotated image datasets remains a major obstacle, as the annotation process is time-consuming and costly. This study aims to overcome this challenge by proposing an automated method for annotating a large database of medical radiology images based on their semantic similarity.
RESULTS: An automated, unsupervised approach is used to create a large annotated dataset of medical radiology images originating from the Clinical Hospital Centre Rijeka, Croatia. The pipeline is built by data-mining three different types of medical data: images, DICOM metadata and narrative diagnoses. The optimal feature extractors are then integrated into a multimodal representation, which is then clustered to create an automated pipeline for labelling a precursor dataset of 1,337,926 medical images into 50 clusters of visually similar images. The quality of the clusters is assessed by examining their homogeneity and mutual information, taking into account the anatomical region and modality representation.
CONCLUSIONS: The results indicate that fusing the embeddings of all three data sources together provides the best results for the task of unsupervised clustering of large-scale medical data and leads to the most concise clusters. Hence, this work marks the initial step towards building a much larger and more fine-grained annotated dataset of medical radiology images.

Proteins are complex biomolecules essential for numerous biological processes, making them crucial targets for advancements in molecular biology, medical research, and drug design. Understanding their intricate, hierarchical structures, and functions is vital for progress in these fields. To capture this complexity, we introduce Multimodal Protein Representation Learning (MPRL), a novel framework for symmetry-preserving multimodal pretraining that learns unified, unsupervised protein representations by integrating primary and tertiary structures. MPRL employs Evolutionary Scale Modeling (ESM-2) for sequence analysis, Variational Graph Auto-Encoders (VGAE) for residue-level graphs, and PointNet Autoencoder (PAE) for 3D point clouds of atoms, each designed to capture the spatial and evolutionary intricacies of proteins while preserving critical symmetries. By leveraging Auto-Fusion to synthesize joint representations from these pretrained models, MPRL ensures robust and comprehensive protein representations. Our extensive evaluation demonstrates that MPRL significantly enhances performance in various tasks such as protein-ligand binding affinity prediction, protein fold classification, enzyme activity identification, and mutation stability prediction. This framework advances the understanding of protein dynamics and facilitates future research in the field. Our source code is publicly available at https://github.com/HySonLab/Protein_Pretrain.

The integration of information from multiple modalities is a highly active area of research. Previous techniques have predominantly focused on fusing shallow features or high-level representations generated by deep unimodal networks, which only capture a subset of the hierarchical relationships across modalities. However, previous methods are often limited to exploiting the fine-grained statistical features inherent in multimodal data. This paper proposes an approach that densely integrates representations by computing image features\' means and standard deviations. The global statistics of features afford a holistic perspective, capturing the overarching distribution and trends inherent in the data, thereby facilitating enhanced comprehension and characterization of multimodal data. We also leverage a Transformer-based fusion encoder to effectively capture global variations in multimodal features. To further enhance the learning process, we incorporate a contrastive loss function that encourages the discovery of shared information across different modalities. To validate the effectiveness of our approach, we conduct experiments on three widely used multimodal sentiment analysis datasets. The results demonstrate the efficacy of our proposed method, achieving significant performance improvements compared to existing approaches.

Action recognition is an important component of human-computer interaction, and multimodal feature representation and learning methods can be used to improve recognition performance due to the interrelation and complementarity between different modalities. However, due to the lack of large-scale labeled samples, the performance of existing ConvNets-based methods are severely constrained. In this paper, a novel and effective multi-modal feature representation and contrastive self-supervised learning framework is proposed to improve the action recognition performance of models and the generalization ability of application scenarios. The proposed recognition framework employs weight sharing between two branches and does not require negative samples, which could effectively learn useful feature representations by using multimodal unlabeled data, e.g., skeleton sequence and inertial measurement unit signal (IMU). The extensive experiments are conducted on two benchmarks: UTD-MHAD and MMAct, and the results show that our proposed recognition framework outperforms both unimodal and multimodal baselines in action retrieval, semi-supervised learning, and zero-shot learning scenarios.

Data processing in robotics is currently challenged by the effective building of multimodal and common representations. Tremendous volumes of raw data are available and their smart management is the core concept of multimodal learning in a new paradigm for data fusion. Although several techniques for building multimodal representations have been proven successful, they have not yet been analyzed and compared in a given production setting. This paper explored three of the most common techniques, (1) the late fusion, (2) the early fusion, and (3) the sketch, and compared them in classification tasks. Our paper explored different types of data (modalities) that could be gathered by sensors serving a wide range of sensor applications. Our experiments were conducted on Amazon Reviews, MovieLens25M, and Movie-Lens1M datasets. Their outcomes allowed us to confirm that the choice of fusion technique for building multimodal representation is crucial to obtain the highest possible model performance resulting from the proper modality combination. Consequently, we designed criteria for choosing this optimal data fusion technique.

How are words connected to the thoughts they help to express? Recent brain imaging studies suggest that word representations are embodied in different neural systems through which the words are experienced. Building on this idea, embodied approaches such as the Concept Attribute Representations (CAR) theory represents concepts as a set of semantic features (attributes) mapped to different brain systems. An intriguing challenge to this theory is that people weigh concept attributes differently based on context, i.e., they construct meaning dynamically according to the combination of concepts that occur in the sentence. This research addresses this challenge through the Context-dEpendent meaning REpresentations in the BRAin (CEREBRA) neural network model. Based on changes in the brain images, CEREBRA quantifies the effect of sentence context on word meanings. Computational experiments demonstrated that words in different contexts have different representations, the changes observed in the concept attributes reveal unique conceptual combinations, and that the new representations are more similar to the other words in the sentence than to the original representations. Behavioral analysis further confirmed that the changes produced by CEREBRA are actionable knowledge that can be used to predict human responses. These experiments constitute a comprehensive evaluation of CEREBRA\'s context-based representations, showing that CARs can be dynamic and change based on context. Thus, CEREBRA is a useful tool for understanding how word meanings are represented in the brain, providing a framework for future interdisciplinary research on the mental lexicon.

Our thoughts can influence sleep, but the underlying mechanisms are unknown. According to the theory of \"embodied cognition,\" the semantic content of cognitive processes is represented by multimodal networks in the brain, which include body-related functions. Such multimodal representations could offer a mechanism, which explains mutual influences between cognition and sleep. Here we tested whether sleep-related words are represented in multimodal networks by examining the effect of congruent versus incongruent body positions on word processing during wakefulness. We experimentally manipulated the body position of 66 subjects (19-40 years old) between standing upright and lying down. Sleep- and activity-related words were presented around the individual speech recognition threshold. Our results show that word processing was facilitated in congruent body positions (sleep words: lying down and activity words: standing upright) compared with incongruent body positions, as indicated by a reduced N400 in the congruent condition with the lowest volume. In addition, early sensory components of the ERP (N180 and P280) were enhanced, suggesting that words were also acoustically better understood in a congruent body position. However, the difference in ERPs did not translate to differences on a behavioral level. Our results support the prediction of embodied processing of sleep- and activity-related words. Body position potentially induces a pre-activation of multimodal networks, thereby enhancing access to the semantic concepts of words related to current body position. The link between semantic meaning and body-related function could be a key element in explaining the influences of cognitive processing on sleep.