Zhenguang Wang, Bo Li, Wenhui Tan, Peng Cao, Yang Wang, Jia Duan, Fei Wang, Osmar Zaiane. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Depression recognition is critical for early detection and treatment. Existing works focus on modeling coarse-grained multimodal representation to estimate the depression level. However, these approaches often overlook the inherent locality of depressive representation, resulting in weak and sparse depressive frames being overlooked. In addition, they neglect the inter modal correlations and intra-modal patterns of mood change, limiting the learning of multimodal complementary information. Therefore, we present a Locality-Aware Multimodal Depression (LAMD) recognition model. Specifically, LAMD contains three innovations: 1) Considering the sparsity of depressive features, we propose an Adaptive Temporal Attention (ATA) module to adaptively highlight keyframes with depressive features and suppress irrelevant frames. Additionally, we introduce Segment Information Sharing (SIS) strategy to overcome the limitation of inter-segment independence, enabling global awareness of depressive features within the whole segment. 2) We revisit the audio-video multimodal interaction from the perspectives of inter-modal correlation and intra-modal smoothness, introducing frame-level multimodal attention consistency constraints and smooth constraints. Furthermore, we propose a local cross attention to enhance the inter-modal interactions in adjacent time. 3) Extensive experiments on several datasets demonstrate that LAMD achieves superior performance, with up to 7.21 RMSE and 76.77% F1-score on the AVEC2014 and NJAD dataset, outperforming the prior art by a notable 0.22% and 1.88% margin, respectively. Moreover, visual analysis reveals that LAMD can adaptively perceive depressive keyframes and focus on fine-grained facial regions known for capturing subtle depressive expressions.

This paper explores a challenging yet realistic scenario: semi-supervised domain generalization (SSDG) that includes label scarcity and domain shift problems. We pinpoint that the limitations of previous SSDG methods lie in 1) neglecting the difference between domain shifts existing within a training dataset (intra-domain shift, IDS) and those occurring between training and testing datasets (cross-domain shift, CDS) and 2) overlooking the interplay between label scarcity and domain shifts, resulting in these methods merely stitching together semi-supervised learning (SSL) and domain generalization (DG) techniques. Considering these limitations, we propose a novel perspective to decompose SSDG into the combination of unsupervised domain adaptation (UDA) and DG problems. To this end, we design a causal augmentation and disentanglement framework (CausalAD) for semi-supervised domain generalized medical image segmentation. Concretely, CausalAD involves two collaborative processes: an augmentation process, which utilizes disentangled style factors to perform style augmentation for UDA, and a disentanglement process, which decouples domain-invariant (content) and domain-variant (noise and style) features for DG. Furthermore, we propose a proxy-based self-paced training strategy (ProSPT) to guide the training of CausalAD by gradually selecting unlabeled image pixels with high-quality pseudo labels in a self-paced training manner. Finally, we introduce a hierarchical structural causal model (HSCM) to explain the intuition and concept behind our method. Extensive experiments in the cross-sequence, cross-site, and cross-modality semi-supervised domain generalized medical image segmentation settings show the effectiveness of CausalAD and its superiority over the state-of-the-art. The code is available at https://github.com/Senyh/CausalAD.

The progression of medical image analysis methodologies has significantly assisted fundus clinical decision-making, such as disease diagnosis and lesion segmentation. However, low-quality fundus images bring a series of challenges to the automatic screening of diseases and the segmentation of lesions. Most existing methods primarily concentrate on enhancing image quality by utilizing the supervision of paired fundus images, which are difficult to collect in real medical applications. High-quality reference images are essential for guiding quality enhancement. To this end, we propose an enhancement method for low-quality fundus images, called RF-IQE, to alleviate the requirement for paired training images and only requires low-quality fundus images. Specifically, we first construct the patch-level high-/low-quality domains by employing a rule-based quality assessment scheme. Then, to achieve the fundus image quality enhancement and unified illumination styles simultaneously, we formulate them as a patch quality domain adaptation and a multi-style domain adaptation, respectively. We qualitatively and quantitatively demonstrate that our reference-free image quality enhancement network outperforms the conventional methods and exhibits comparable performance than the deep learning-based image enhancement methods with paired images on both the EyeQ and Messidor datasets. Furthermore, we also investigate the influence of the RF-IQE method on various fundus imaging analysis tasks, including vessel segmentation, optic disc segmentation, lesion segmentation, and disease classification.

Brain functional connectivity analysis is important for understanding brain development and brain disorders. Recent studies have suggested that the variations of functional connectivity among multiple subnetworks are closely related to the development of diseases. However, the existing works failed to sufficiently capture the complex correlation patterns among the subnetworks and ignored the learning of heterogeneous structural information across the subnetworks. To address these issues, we formulate a new paradigm for constructing and analyzing high-order heterogeneous functional brain networks via meta-paths and propose a Heterogeneous Graph representation Learning framework (BrainHGL). Our framework consists of three key aspects: 1) Meta-path encoding for capturing rich heterogeneous topological information, 2) Meta-path interaction for exploiting complex association patterns among subnetworks and 3) Meta-path aggregation for better meta-path fusion. To the best of our knowledge, we are the first to formulate the heterogeneous brain networks for better exploiting the relationship between the subnetwork interactions and the mental disease We evaluate BrainHGL on the private center Nanjing Medical University dataset (center NMU) and the public Autism Brain Imaging Data Exchange (ABIDE) dataset. We demonstrate the effectiveness of the proposed model across various disease classification tasks, including major depression disorder (MDD), bipolar disorder (BD) and autism spectrum disorder (ASD) diagnoses. In addition, our model provides deeper insights into disease interpretability, including the critical brain subnetwork connectivities, brain regions and functional pathways. We also identified disease subtypes consistent with previous neuroscientific studies by our model, which benefits the disease identification performance. The code is available at https://github.com/IntelliDAL/Graph/BrainHGL.

Medical images acquired under suboptimal conditions often suffer from quality degradation, such as low-light, blurring, and artifacts. Such degradations obscure the lesions and anatomical structures in medical images, making it difficult to distinguish key pathological regions. This significantly increases the risk of misdiagnosis by automated medical diagnostic systems or clinicians. To address this challenge, we propose a multi-Color space-based quality enhancement network (MSQNet) that effectively eliminates global low-quality factors while preserving pathology-related characteristics for improved clinical observation and analysis. We first revisit the properties of image quality enhancement in different color spaces, where the V-channel in the HSV space can better represent the contrast and brightness enhancement process, whereas the A/B-channel in the LAB space is more focused on the color change of low-quality images. The proposed framework harnesses the unique properties of different color spaces to optimize the image enhancement process. Specifically, we propose a pathology-preserving transformer, designed to selectively aggregate features across different color spaces and enable comprehensive multiscale feature fusion. Leveraging these capabilities, MSQNet effectively enhances low-quality RGB medical images while preserving key pathological features, thereby establishing a new paradigm in medical image enhancement. Extensive experiments on three public medical image datasets demonstrate that MSQNet outperforms traditional enhancement techniques and state-of-the-art methods, in terms of both quantitative metrics and qualitative visual assessment. MSQNet successfully improves image quality while preserving pathological features and anatomical structures, facilitating accurate diagnosis and analysis by medical professionals and automated systems.

Graph similarity estimation is a challenging task due to the complex graph structures. Though important and well-studied, three critical aspects are yet to be fully handled in a unified framework: 1) how to learn richer cross-graph interactions from a pairwise node perspective; 2) how to map the similarity matrix into a similarity score by exploiting the inherent structure in the similarity matrix; and 3) how to establish a self-supervised learning mechanism for graph similarity learning. To solve these issues, we explore multiple attention and self-supervised mechanisms for graph similarity learning in this work. More specifically, we propose a unified self-supervised nodewise attention-guided graph similarity learning framework (SNA-GSL) involving: 1) a correlation-guided contrastive learning for capturing valuable node embeddings and 2) a graph similarity learning for predicting similarity scores with multiple proposed attention mechanisms. Extensive experimental results on graph-graph regression task and graph classification task demonstrate that the proposed SNA-GSL performs favorably against state-of-the-art methods. Moreover, the remarkable achievement of our model in the graph classification task is a clear indication of its exceptional generalization capabilities. The code is available at https://github.com/IntelliDAL/Graph/SNA-GSL.

Text style transfer, an important research direction in natural language processing, aims to adapt the text to various preferences but often faces challenges with limited resources. In this work, we introduce a novel method termed Style Extraction and Tunable Inference via Dual-level Transferable Prompt Learning (SETTP) for effective style transfer in low-resource scenarios. First, SETTP learns source style-level prompts containing fundamental style characteristics from high-resource style transfer. During training, the source style-level prompts are transferred through an attention module to derive a target style-level prompt for beneficial knowledge provision in low-resource style transfer. Additionally, we propose instance-level prompts obtained by clustering the target resources based on the semantic content to reduce semantic bias. We also propose an automated evaluation approach of style similarity based on alignment with human evaluations using ChatGPT-4. Our experiments across three resourceful styles show that SETTP requires only 1/20th of the data volume to achieve performance comparable to state-of-the-art methods. In tasks involving scarce data like writing style and role style, SETTP outperforms previous methods by 16.24%.

Text style transfer (TST) is crucial in natural language processing, aiming to endow text with a new style without altering its meaning.In real-world scenarios, not all styles have abundant resources.This work introduces TWIST (reusing Transferable Weight Increments for Style Text generation), a novel framework to mitigate data scarcity by utilizing style features in weight increments to transfer low-resource styles effectively.During target style learning, we derive knowledge via a specially designed weight pool and initialize the parameters for the unseen style.To enhance the effectiveness of merging, the target style weight increments are often merged from multiple source style weight increments through singular vectors.Considering the diversity of styles, we also designed a multi-key memory network that simultaneously focuses on task-and instance-level information to derive the most relevant weight increments.Results from multiple style transfer datasets show that TWIST demonstrates remarkable performance across different backbones, achieving particularly effective results in low-resource scenarios.

The dynamic brain network learning methods ignored the separation of redundant disease-irrelevant information, resulting in the model only achieving suboptimal diagnosis results. Meanwhile, the supervised learning scheme inevitably suffers from poor generalization due to the limited data. To address these problems, we propose a Self-supervised Dynamic Brain network Disentangled representation learning framework named SDBD, which incorporates 1) a dynamic topology-aware encoder for capturing diverse topological information, 2) a cross decoder for reconstructing the graph structure and 3) a spatio-temporal learning model based on the multi-head self-attention mechanism for classification. To disentangle the disease-related information from the dynamic brain networks, we design a temporal contrastive loss and a structure reconstruction loss. We evaluate our model on three real-world mental diseases: Autism Spectrum Disorder (ASD), Major Depressive Disorder (MDD), and Bipolar Disorder (BD). The results indicate significant improvements in our SDBD over the state-of-the-art methods owing to the disentangled disease-related information. Moreover, our method can identify the biomarkers associated with the diseases, which is consistent with the previous studies. To the best of our knowledge, our work is the first attempt to disentangle the disease-related information for the dynamic brain network analysis. The code is available at https://github.com/IntelliDAL/Graph/tree/main/SDBD.

\beginabstract Dynamic graphs play an important role in many fields like social relationship analysis, recommender systems and medical science, as graphs evolve over time. It is fundamental to capture the evolution patterns for dynamic graphs. Existing works mostly focus on constraining the temporal smoothness between neighbor snapshots, however, fail to capture sharp shifts, which can be beneficial for graph dynamics embedding. To solve it, we assume the evolution of dynamic graph nodes can be split into temporal shift embedding and temporal consistency embedding. Thus, we propose the Self-supervised Temporal-aware Dynamic Graph representation Learning framework (STDGL) for disentangling the temporal shift embedding from temporal consistency embedding via a well-designed auxiliary task from the perspectives of both node local and global connectivity modeling in a self-supervised manner, further enhancing the learning of interpretable graph representations and improving the performance of various downstream tasks. Extensive experiments on link prediction, edge classification and node classification tasks demonstrate STDGL successfully learns the disentangled temporal shift and consistency representations. Furthermore, the results indicate significant improvements in our STDGL over the state-of-the-art methods, and appealing interpretability and transferability owing to the disentangled node representations. \endabstract

Medical image analysis techniques have been employed in diagnosing and screening clinical diseases. However, both poor medical image quality and illumination style inconsistency increase uncertainty in clinical decision-making, potentially resulting in clinician misdiagnosis. The majority of current image enhancement methods primarily concentrate on enhancing medical image quality by leveraging high-quality reference images, which are challenging to collect in clinical applications. In this study, we address image quality enhancement within a fully self-supervised learning setting, wherein neither high-quality images nor paired images are required. To achieve this goal, we investigate the potential of self-supervised learning combined with domain adaptation to enhance the quality of medical images without the guidance of high-quality medical images. We design a Domain Adaptation Self-supervised Quality Enhancement framework, called DASQE. More specifically, we establish multiple domains at the patch level through a designed rule-based quality assessment scheme and style clustering. To achieve image quality enhancement and maintain style consistency, we formulate the image quality enhancement as a collaborative self-supervised domain adaptation task for disentangling the low-quality factors, medical image content, and illumination style characteristics by exploring intrinsic supervision in the low-quality medical images. Finally, we perform extensive experiments on six benchmark datasets of medical images, and the experimental results demonstrate that DASQE attains state-of-the-art performance. Furthermore, we explore the impact of the proposed method on various clinical tasks, such as retinal fundus vessel/lesion segmentation, nerve fiber segmentation, polyp segmentation, skin lesion segmentation, and disease classification. The results demonstrate that DASQE is advantageous for diverse downstream image analysis tasks.

Consistency regularization and pseudo labeling-based semi-supervised methods perform co-training using the pseudo labels from multi-view inputs. However, such co-training models tend to converge early to a consensus, degenerating to the self-training ones, and produce low-confidence pseudo labels from the perturbed inputs during training. To address these issues, we propose an Uncertainty-guided Collaborative Mean-Teacher (UCMT) for semi-supervised semantic segmentation with the high-confidence pseudo labels. Concretely, UCMT consists of two main components: 1) collaborative mean-teacher (CMT) for encouraging model disagreement and performing co-training between the sub-networks, and 2) uncertainty-guided region mix (UMIX) for manipulating the input images according to the uncertainty maps of CMT and facilitating CMT to produce high-confidence pseudo labels. Combining the strengths of UMIX with CMT, UCMT can retain model disagreement and enhance the quality of pseudo labels for the co-training segmentation. Extensive experiments on four public medical image datasets including 2D and 3D modalities demonstrate the superiority of UCMT over the state-of-the-art. Code is available at: https://github.com/Senyh/UCMT.

Retinal fundus images have been applied for the diagnosis and screening of eye diseases, such as Diabetic Retinopathy (DR) or Diabetic Macular Edema (DME). However, both low-quality fundus images and style inconsistency potentially increase uncertainty in the diagnosis of fundus disease and even lead to misdiagnosis by ophthalmologists. Most of the existing fundus image enhancement methods mainly focus on improving the image quality by leveraging the guidance of high-quality images, which is difficult to be collected in medical applications. In this paper, we tackle image quality enhancement in a fully unsupervised setting, i.e., neither paired images nor high-quality images. To this end, we explore the potential of the self-supervised task for improving the quality of fundus images without the requirement of high-quality reference images, and proposed a Domain Adaptation Self-supervised Quality Enhancement framework, named DASQE. Specifically, we construct multiple patch-wise domains via a well-designed rule-based quality assessment scheme and style clustering. To achieve robust low-quality image enhancement and address style inconsistency, we formulate two self-supervised domain adaptation tasks to disentangle the features of image content, low-quality factors and style information by exploring intrinsic supervision signals within the low-quality images. Extensive experiments are conducted on four benchmark datasets, and results show that our DASQE method achieves new state-of-the-art performance when only low-quality images are available.

Diabetic retinopathy (DR) is one of the most serious complications of diabetes and is a prominent cause of permanent blindness. However, the low-quality fundus images increase the uncertainty of clinical diagnosis, resulting in a significant decrease on the grading performance of the fundus images. Therefore, enhancing the image quality is essential for predicting the grade level in DR diagnosis. In essence, we are faced with three challenges: (I) How to appropriately evaluate the quality of fundus images; (II) How to effectively enhance low-quality fundus images for providing reliable fundus images to ophthalmologists or automated analysis systems; (III) How to jointly train the quality assessment and enhancement for improving the DR grading performance. Considering the importance of image quality assessment and enhancement for DR grading, we propose a collaborative learning framework to jointly train the subnetworks of the image quality assessment as well as enhancement, and DR disease grading in a unified framework. The key contribution of the proposed framework lies in modelling the potential correlation of these tasks and the joint training of these subnetworks, which significantly improves the fundus image quality and DR grading performance. Our framework is a general learning model, which may be useful in other medical images with low-quality data. Extensive experimental results have shown that our method outperforms state-of-the-art DR grading methods by a considerable 73.6% ACC/71.2% Kappa and 88.5% ACC/86.3% Kappa on Messidor and EyeQ benchmark datasets, respectively. In addition, our method significantly enhances the low-quality fundus images while preserving fundus structure features and lesion information. To make the framework more general, we also evaluate the enhancement results in more downstream tasks, such as vessel segmentation.

The less training data and insufficient supervision limit the performance of the deep supervised models for brain disease diagnosis. It is significant to construct a learning framework that can capture more information in limited data and insufficient supervision. To address these issues, we focus on self-supervised learning and aim to generalize the self-supervised learning to the brain networks, which are non-Euclidean graph data. More specifically, we propose an ensemble masked graph self-supervised framework named BrainGSLs, which incorporates 1) a local topological-aware encoder that takes the partially visible nodes as input and learns these latent representations, 2) a node-edge bi-decoder that reconstructs the masked edges by the representations of both the masked and visible nodes, 3) a signal representation learning module for capturing temporal representations from BOLD signals and 4) a classifier used for the classification. We evaluate our model on three real medical clinical applications: diagnosis of Autism Spectrum Disorder (ASD), diagnosis of Bipolar Disorder (BD) and diagnosis of Major Depressive Disorder (MDD). The results suggest that the proposed self-supervised training has led to remarkable improvement and outperforms state-of-the-art methods. Moreover, our method is able to identify the biomarkers associated with the diseases, which is consistent with the previous studies. We also explore the correlation of these three diseases and find the strong association between ASD and BD. To the best of our knowledge, our work is the first attempt of applying the idea of self-supervised learning with masked autoencoder on the brain network analysis.

Most recent semantic segmentation methods adopt a U-Net framework with an encoder-decoder architecture. It is still challenging for U-Net with a simple skip connection scheme to model the global multi-scale context: 1) Not each skip connection setting is effective due to the issue of incompatible feature sets of encoder and decoder stage, even some skip connection negatively influence the segmentation performance; 2) The original U-Net is worse than the one without any skip connection on some datasets. Based on our findings, we propose a new segmentation framework, named UCTransNet (with a proposed CTrans module in U-Net), from the channel perspective with attention mechanism. Specifically, the CTrans (Channel Transformer) module is an alternate of the U-Net skip connections, which consists of a sub-module to conduct the multi-scale Channel Cross fusion with Transformer (named CCT) and a sub-module Channel-wise Cross-Attention (named CCA) to guide the fused multi-scale channel-wise information to effectively connect to the decoder features for eliminating the ambiguity. Hence, the proposed connection consisting of the CCT and CCA is able to replace the original skip connection to solve the semantic gaps for an accurate automatic medical image segmentation. The experimental results suggest that our UCTransNet produces more precise segmentation performance and achieves consistent improvements over the state-of-the-art for semantic segmentation across different datasets and conventional architectures involving transformer or U-shaped framework. Code: https://github.com/McGregorWwww/UCTransNet.

Abstract Live Streaming Commerce (LSC) is proliferating in China and gaining traction worldwide. LSC is an e-commerce service where sellers communicate with consumers through live streaming while consumers can place orders within the same system. Despite the significant involvement of consumers in LSC, it has not been systematically analyzed how consumers make shopping decisions when engaging with LSC. In this paper, we conduct a mixed-methods study, consisting of surveys ( N 1 = 240) and follow-up interviews ( N 2 = 16) with LSC consumers. We focus on two features of LSC, i.e., the communication between merchants and consumers through live streaming and the participation of streamers, and aim to understand how these changes influence consumers’ decision-making process in LSC. We find that LSC enables merchants to exchange information with consumers based on their needs and provide additional customer services. Because of the appropriate information about the products they acquire and the enjoyable shopping atmosphere, consumers are willing to purchase products in LSC. As the intermediaries between merchants and consumers, streamers utilize their independent identity from merchants to enhance consumers’ awareness of shopping and persuade their online shopping decisions. Moreover, we consider the opportunities and challenges of current LSC services and provide implications for LSC services and the research community regarding the development of LSC.

Alzheimer’s disease (AD) is known as one of the major causes of dementia and is characterized by slow progression over several years. There have been efforts to identify the risk of developing AD in its earliest time. Recently, multi-task feature learning (MTFL) methods with sparsity-inducing $\ell_{2,1}$-norm have been widely studied to select a discriminative feature subset from MRI features. However, they ignore the complex relationships among imaging markers and among cognitive outcomes. Constructing the relationships with simple Pearson correlation coefficient may degrade model generalizability. To better capture the complicated but more flexible relationship between the cognitive scores and the neuroimaging measures, we propose a two-stage framework to jointly learn the structure within the feature correlation as well as within the task correlation. Moreover, we propose a dual graph regularization to encode the learned correlation structure. It is able to guide the training procedure of MTFL by incorporating both the inherent correlations. Extensive results on benchmark datasets show that for the proposed FTSMTFL model trained with the dual graph regularization, the proposed joint training framework outperforms existing methods and achieves state-of-the-art cognitive prediction performance of AD.

Modeling spatio-temporal dynamics in functional brain networks is critical for underlying the functional mechanism of autism spectrum disorder (ASD). In our study, we propose an end-to-end framework called temporal graph representation learning for brain networks, which thoroughly captures spatio-temporal features in resting-state functional magnetic resonance imaging (rs-fMRI) data. Specifically, we first transform rs-fMRI time-series into temporal multi-graph using a sliding window technique. A temporal multi-graph clustering is then designed to eliminate the inconsistency of the temporal multi-graph series. Then, a graph structure aware LSTM (GSA-LSTM) is proposed to capture the spatio-temporal embedding for temporal graphs. The proposed GSA-LSTM can not only capture discriminative features for prediction but also impute the incomplete graphs for the temporal multi-graph series. Extensive experiments on autism brain imaging data exchange (ABIDE) dataset shows the effectiveness of our proposed framework. The results demonstrate that the proposed dynamic brain network embedding learning outperforms the state of-the-art brain network classification models. Furthermore, the obtained clustering results are consistent with the previous neuroimaging-derived evidence of biomarkers for autism spectrum disorder (ASD).

MRI brain segmentation plays an important part in computer-aided diagnosis, which visually reveals the changes in brain structure for doctors to quickly and accurately discover and treat diseases related to brain tissue morphology. The fuzzy C-means (FCM) algorithm performs well when the segmenting images with no noise and with intensity uniformity. However, the MRI brain images are always defective in noise and intensity nonuniformity and thus we propose a novel FCM algorithm named adaptive FCM with neighborhood membership (FCM_anm). We design a filtering process with neighborhood membership to reduce the negative influence of noise and a novel objective function which further considers the spatial membership information adaptively. Finally, to verify the performance of our method, several experiments comparing among the Experimental results demonstrate the proposed method consistently outperforms the state-of-the-art FCM-based algorithms in synthetic images, simulated and real brain MR images with effects of the noise and intensity non-uniformity.

Diabetic retinopathy (DR) is one of the most concerning, common and serious diseases in the ophthalmology community. Early detection and treatment of DR can significantly reduce the risk of vision loss in patients. Traditional DR automatic classification algorithms rely on the precise detection of microaneurysms (MA) and hemorrhage (H) lesions. Such lesion annotation is an expensive and time-consuming process, hence it is expected to develop automatic grading methods with only image-level annotations. The lack of the position of MA and H hinders the traditional supervised algorithms for the accurate identification. In our work, we formulate the weakly supervised DR grading as a multi-instance learning problem, and propose a domain adaptation multi-instance learning with attention mechanism for DR grading. Specifically, labeled instances are generated by cross-domain to filter irrelevant instances in the target domain. To model the relationship between the suspicious instances and bag label, a multi-instance learning with attention mechanism is developed to acquire the location information of highly suspected lesions and predict the grade of DR. We evaluate our proposed algorithm on the Messidor dataset, and the experimental results demonstrate that it achieves an average accuracy of 0.764 and an AUC value of 0.749 respectively, outperforming state-of-the-art approaches.

Skin conditions affect 1.9 billion people. Because of a shortage of dermatologists, most cases are seen instead by general practitioners with lower diagnostic accuracy. Current skin disease researches adopt the auto-classification system for improving the accuracy rate of skin disease classification. It is therefore an important task to develop Computer Aided Detection (CAD) systems that can aid/enhance dermatologists workflow and improve the classification performances. However, the long-tailed class distribution in the database and the limitation of ability to achieve a spatially invariant features make this problem challenging. We propose a ST-MetaDiagnosis, which utilizes meta-learning and spatial transform learning to facilitate quick adaptation and generalization of deep neural networks trained on the common diseases data for identification of rare diseases with much less annotated data. In particular, in order to predict the target risk where there are limited data samples, we train a meta-learner with spatial transforming from a set of related risk prediction tasks which learns how a good predictor is learned. The meta-learned can be directly used in target risk prediction, and the limited available samples can be used for further fine-tuning the model performance. Experiments on the recent ISIC 2018 skin lesion classification dataset show that our ST-MetaDiagnosis obtains 64.6% (accuracy) and 64.4% (F1-score) on the diagnosis of actinic keratosis, vascular lesion and dermatofibroma, demonstrating that ST-MetaDiagnosis can improve performance for predicting target risk with low resources comparing with the predictor trained on the limited samples available for this risk.

Previous work has shown that machine learning algorithms lend themselves to clinical decision-making and are a valuable tool for physicians. For clinical data, it is often necessary to assign multiple labels to a patient record by choosing from a large number of potential labels. A key problem in learning from multi-labelled data is how to exploit the information contained in the correlations between labels. The hypergraph-based multi-label learning method learns from data by exploiting the spectral property of the hypergraph that encodes the correlation structure of labels. However, the problem with this method is the difficulty with which interpretations can be made. This is mainly due to its inability to recognize the importance of key features in the original feature space. Moreover, it is hard to comprehensively capture the complex structure of the correlations between labels. To overcome these difficulties and improve interpretability, we propose an l21-norm regularized Graph Laplacian multi-label learning to perform feature selection and label embedding simultaneously. In-depth experimental studies, using the publicly available Medical Information Mart for Intensive Care (MIMIC-III) database, validate the effectiveness of our approach.

Machine learning algorithms and multivariate data analysis methods have been widely utilized in the field of Alzheimer's disease (AD) research in recent years. Predicting cognitive performance of subjects from neuroimage measures and identifying relevant imaging biomarkers are important research topics in the study of Alzheimer's disease. Multi-task based feature learning (MTFL) have been widely studied to select a discriminative feature subset from MRI features, and improve the performance by incorporating inherent correlations among multiple clinical cognitive measures. It is known that the brain imaging measures are often correlated with each other, and AD is closely related to the inter-correlation among different brain regions. However, the multi-task based feature learning (MTFL) method neglects the inherent correlation among brain imaging measures. We present a novel regularized multi-task learning approach via a joint sparsity-inducing regularization to effectively incorporate both a relatedness among multiple cognitive score prediction tasks and a useful inherent correlation between brain imaging measures by exploiting correlations among features. It allows the simultaneous selection of a common set of biomarkers for all tasks and the preservation of the inherent structure of imaging measures. The reported experiments on the ADNI dataset show that the proposed method is effective and promising.

Alzheimer’s disease (AD), the most common type of dementia, not only imposes a huge financial burden on the health care system, but also a psychological and emotional burden on patients and their families. There is thus an urgent need to infer trajectories of cognitive performance over time and identify biomarkers predictive of the progression. In this article, we propose the multi-task learning with fused Laplacian sparse group lasso model, which can identify biomarkers closely related to cognitive measures due to its sparsity-inducing property, and model the disease progression with a general weighted (undirected) dependency graphs among the tasks. An efficient alternative directions method of multipliers based optimization algorithm is derived to solve the proposed non-smooth objective formulation. The effectiveness of the proposed model is demonstrated by its superior prediction performance over multiple state-of-the-art methods and accurate identification of compact sets of cognition-relevant imaging biomarkers that are consistent with prior medical studies.

Class imbalance is one of the challenging problems for machine learning in many real-world applications. Other issues, such as within-class imbalance and high dimensionality, can exacerbate the problem. We propose a method HPS-DRS that combines two i

Many lung nodule computer-aided detection methods have been proposed to help radiologists in their decision making. Because high sensitivity is essential in the candidate identification stage, there are countless false positives produced by the initial suspect nodule generation process, giving more work to radiologists. The difficulty of false positive reduction lies in the variation of the appearances of the potential nodules, and the imbalance distribution between the amount of nodule and non-nodule candidates in the dataset. To solve these challenges, we extend the random subspace method to a novel Cost Sensitive Adaptive Random Subspace ensemble (CSARS), so as to increase the diversity among the components and overcome imbalanced data classification. Experimental results show the effectiveness of the proposed method in terms of G-mean and AUC in comparison with commonly used methods.

The performance of traditional classification algorithms can be limited on imbalanced datasets. In recent years, the imbalanced data learning problem has drawn significant interest. In this work, we focus on designing modifications to neural network, in order to appropriately tackle the problem of multiclass imbalance. We propose a hybrid method that combines two ideas: diverse random subspace ensemble learning with evolutionary search, to improve the performance of neural network on multiclass imbalanced data. An evolutionary search technique is utilized to optimize the misclassification cost under the guidance of imbalanced data measures. Moreover, the diverse random subspace ensemble employs the minimum overlapping mechanism to provide diversity so as to improve the performance of the learning and optimization of neural network. We have demonstrated experimentally using UCI datasets that our approach can achieve better result than state-of-the-art methods for imbalanced data.

Traditional classification algorithms can be limited in their performance on imbalanced datasets. In recent years, the imbalanced data learning problem has drawn significant interest. In this work, we focus on designing modifications to neural network, in order to appropriately tackle the problem of multiclass imbalance. We propose a method that combines two ideas: diverse random subspace ensemble learning with evolutionary search, to improve the performance of neural network on multiclass imbalanced data. An evolutionary search technique is utilized to optimize the misclassification cost under the guidance of imbalanced data measures. Moreover, the diverse random subspace ensemble employs the minimum overlapping mechanism to provide diversity so as to improve the performance of the learning and optimization of neural network. Furthermore, the ensemble framework can determine the optimal amount of non-redundant components automatically. We have demonstrated experimentally using UCI datasets that our approach can achieve significantly better result than state-of-the-art methods for imbalanced data.

Class imbalance is one of the challenging problems for machine learning in many real-world applications. Many methods have been proposed to address and attempt to solve the problem, including re-sampling and cost-sensitive learning. However, the existing methods have room for improvement since the potentially optimal values of the factors associated with best performance are unknown. Moreover most methods only focus on the binary class imbalance problem, thus there is no efficient solution in multi-class imbalanced learning. This paper presents an effective wrapper framework incorporating the evaluation measure into the objective function of cost sensitive learning as well as re-sampling directly, so as to improve the original methods through optimizing factors influencing the performance on the imbalanced data classification. Comprehensive experimental results on various standard benchmark datasets with different ratios of imbalance show that the influence of optimizing parameters on the solutions for learning imbalanced data is critical, and demonstrate the effectiveness of measure-optimized scheme on the imbalanced data learning.