Depression is a severe psychological condition that affects millions of people worldwide. As depression has received more attention in recent years, it has become imperative to develop automatic methods for detecting depression. Although numerous machine learning methods have been proposed for estimating the levels of depression via audio, visual, and audiovisual emotion sensing, several challenges still exist. For example, it is difficult to extract long-term temporal context information from long sequences of audio and visual data, and it is also difficult to select and fuse useful multi-modal information or features effectively. In addition, how to include other information or tasks to enhance the estimation accuracy is also one of the challenges. In this study, we propose a multi-modal adaptive fusion transformer network for estimating the levels of depression. Transformer-based models have achieved state-of-the-art performance in language understanding and sequence modeling. Thus, the proposed transformer-based network is utilized to extract long-term temporal context information from uni-modal audio and visual data in our work. This is the first transformer-based approach for depression detection. We also propose an adaptive fusion method for adaptively fusing useful multi-modal features. Furthermore, inspired by current multi-task learning work, we also incorporate an auxiliary task (depression classification) to enhance the main task of depression level regression (estimation). The effectiveness of the proposed method has been validated on a public dataset (AVEC 2019 Detecting Depression with AI Sub-challenge) in terms of the PHQ-8 scores. Experimental results indicate that the proposed method achieves better performance compared with currently state-of-the-art methods. Our proposed method achieves a concordance correlation coefficient (CCC) of 0.733 on AVEC 2019 which is 6.2% higher than the accuracy (CCC = 0.696) of the state-of-the-art method.

Sentiment analysis is an important research field aiming to extract and fuse sentimental information from human utterances. Due to the diversity of human sentiment, analyzing from multiple modalities is usually more accurate than from a single modality. To complement the information between related modalities, one effective approach is performing cross-modality interactions. Recently, Transformer-based frameworks have shown a strong ability to capture long-range dependencies, leading to the introduction of several Transformer-based approaches for multimodal processing. However, due to the built-in attention mechanism of the Transformers, only two modalities can be engaged at once. As a result, the complementary information flow in these Transformer-based techniques is partial and constrained. To mitigate this, we propose, TensorFormer, a tensor-based multimodal Transformer framework that takes into account all relevant modalities for interactions. More precisely, we first construct a tensor utilizing the features extracted from each modality, assuming one modality is the target while the remaining tensors serve as the sources. We can generate the corresponding interacted features by calculating source-target attention. This strategy interacts with all involved modalities and generates complementing global information. Experiments on multimodal sentiment analysis benchmark datasets demonstrated the effectiveness of TensorFormer. In addition, we also evaluate TensorFormer in another related area: depression detection and the results reveal significant improvements when compared to other state-of-the-art methods.

Organ segmentation from medical images is one of the most important pre-processing steps in computer-aided diagnosis, but it is a challenging task because of limited annotated data, low-contrast and non-homogenous textures. Compared with natural images, organs in the medical images have obvious anatomical prior knowledge (e.g., organ shape and position), which can be used to improve the segmentation accuracy. In this paper, we propose a novel segmentation framework which integrates the medical image anatomical prior through loss into the deep learning models. The proposed prior loss function is based on probabilistic atlas, which is called as deep atlas prior (DAP). It includes prior location and shape information of organs, which are important prior information for accurate organ segmentation. Further, we combine the proposed deep atlas prior loss with the conventional likelihood losses such as Dice loss and focal loss into an adaptive Bayesian loss in a Bayesian framework, which consists of a prior and a likelihood. The adaptive Bayesian loss dynamically adjusts the ratio of the DAP loss and the likelihood loss in the training epoch for better learning. The proposed loss function is universal and can be combined with a wide variety of existing deep segmentation models to further enhance their performance. We verify the significance of our proposed framework with some state-of-the-art models, including fully-supervised and semi-supervised segmentation models on a public dataset (ISBI LiTS 2017 Challenge) for liver segmentation and a private dataset for spleen segmentation.

Domain generalization (DG) aims to generalize the knowledge learned from multiple source domains to unseen target domains. Existing DG techniques can be subsumed under two broad categories, i.e., domain-invariant representation learning and domain manipulation. Nevertheless, it is extremely difficult to explicitly augment or generate the unseen target data. And when source domain variety increases, developing a domain-invariant model by simply aligning more domain-specific information becomes more challenging. In this paper, we propose a simple yet effective method for domain generalization, named Knowledge Distillation based Domain-invariant Representation Learning (KDDRL), that learns domain-invariant representation while encouraging the model to maintain domain-specific features, which recently turned out to be effective for domain generalization. To this end, our method incorporates multiple auxiliary student models and one student leader model to perform a two-stage distillation. In the first-stage distillation, each domain-specific auxiliary student treats the ensemble of other auxiliary students' predictions as a target, which helps to excavate the domain-invariant representation. Also, we present an error removal module to prevent the transfer of faulty information by eliminating incorrect predictions compared to the true labels. In the second-stage distillation, the student leader model with domain-specific features combines the domain-invariant representation learned from the group of auxiliary students to make the final prediction. Extensive experiments and in-depth analysis on popular DG benchmark datasets demonstrate that our KDDRL significantly outperforms the current state-of-the-art methods.