A5055917015- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Medical Image Segmentation Techniques
- Multimodal Machine Learning Applications
- Video Surveillance and Tracking Methods
- Radiomics and Machine Learning in Medical Imaging
- Medical Imaging and Analysis
- Human Pose and Action Recognition
- COVID-19 diagnosis using AI
- Cancer-related molecular mechanisms research
- Face recognition and analysis
- Machine Learning and Data Classification
- Advanced Image and Video Retrieval Techniques
- AI in cancer detection
- Face and Expression Recognition
- Gait Recognition and Analysis
- Brain Tumor Detection and Classification
- Text and Document Classification Technologies
- Image Retrieval and Classification Techniques
- Anomaly Detection Techniques and Applications
- Machine Learning and ELM
- Sparse and Compressive Sensing Techniques
- Speech Recognition and Synthesis
- Image Processing Techniques and Applications
- Generative Adversarial Networks and Image Synthesis
Nanjing University
2016-2025
Nanjing Drum Tower Hospital
2025
Nanjing University of Science and Technology
2019-2023
University of Wollongong
2021
University of North Carolina at Chapel Hill
2011-2015
University of North Carolina Health Care
2015
Shanghai Medical College of Fudan University
2011
Self-training via pseudo labeling is a conventional, simple, and popular pipeline to leverage unlabeled data. In this work, we first construct strong baseline of self-training (namely ST) for semi-supervised semantic segmentation injecting data augmentations (SDA) on images alleviate overfitting noisy labels as well decouple similar predictions between the teacher student. With simple mechanism, our ST outperforms all existing methods without any bells whistles, e.g., iterative retraining....
Magnetic resonance (MR) imaging is a widely used medical protocol that can be configured to provide different contrasts between the tissues in human body. By setting scanning parameters, each MR modality reflects unique visual characteristic of scanned body part, benefiting subsequent analysis from multiple perspectives. To utilize complementary information modalities, cross-modality image synthesis has aroused increasing research interest recently. However, most existing methods only focus...
In this work, we revisit the weak-to-strong consistency framework, popularized by FixMatch from semi-supervised classification, where prediction of a weakly perturbed image serves as supervision for its strongly version. Intriguingly, observe that such simple pipeline already achieves competitive results against recent advanced works, when transferred to our segmentation scenario. Its success heavily relies on manual design strong data augmentations, however, which may be limited and...
Unsupervised cross-domain person re-identification (Re-ID) faces two key issues. One is the data distribution discrepancy between source and target domains, other lack of discriminative information in domain. From perspective representation learning, this paper proposes a novel end-to-end deep domain adaptation framework to address them. For first issue, we highlight presence camera-level sub-domains as unique characteristic Re-ID, develop “camera-aware” method via adversarial learning. With...
In semi-supervised medical image segmentation, most previous works draw on the common assumption that higher entropy means uncertainty. this paper, we investigate a novel method of estimating We observe that, when assigned different misclassification costs in certain degree, if segmentation result pixel becomes inconsistent, shows relative uncertainty its segmentation. Therefore, present new model, namely, conservative-radical network ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML"...
Few-shot segmentation (FSS) aims to segment unseen classes given only a few annotated samples. Existing methods suffer the problem of feature undermining, i.e., potential novel are treated as background during training phase. Our method alleviate this and enhance embedding on latent classes. In our work, we propose joint-training framework. Based conventional episodic support-query pairs, introduce an additional mining branch that exploits via transferable sub-clusters, new rectification...
Due to the unpredictable location, fuzzy texture, and diverse shape, accurate segmentation of kidney tumor in CT images is an important yet challenging task. To this end, we, paper, present a cascaded trainable model termed as Crossbar-Net. Our method combines two novel schemes: 1) we originally proposed crossbar patches, which consists orthogonal non-squared patches (i.e., vertical patch horizontal patch). The are able capture both global local appearance information tumors from directions...
Accurate segmentation of pelvic organs (i.e., prostate, bladder, and rectum) from CT image is crucial for effective prostate cancer radiotherapy. However, it a challenging task due to: 1) low soft tissue contrast in images 2) large shape appearance variations organs. In this paper, we employ two-stage deep learning-based method, with novel distinctive curve-guided fully convolutional network (FCN), to solve the aforementioned challenges. Specifically, first stage fast robust organ detection...
To address the annotation scarcity issue in some cases of semantic segmentation, there have been a few attempts to develop segmentation model few-shot learning paradigm. However, most existing methods only focus on traditional 1-way setting (i.e., one image contains single object). This is far away from practical tasks where K-way (K > 1) usually required by performing accurate multi-object segmentation. deal with this issue, we formulate task as learning-based pixel classification...
Unsupervised domain adaptation (UDA) methods have shown their promising performance in the cross-modality medical image segmentation tasks. These typical usually utilize a translation network to transform images from source target or train pixel-level classifier merely using translated and original images. However, when there exists large shift between domains, we argue that this asymmetric structure, some extent, could not fully eliminate gap. In paper, present novel deep symmetric...
For medical image segmentation, imagine if a model was only trained using MR images in source domain, how about its performance to directly segment CT target domain? This setting, namely generalizable cross-modality owning clinical potential, is much more challenging than other related settings, e.g., domain adaptation. To achieve this goal, we paper propose novel dual-normalization by leveraging the augmented source-similar and source-dissimilar during our segmentation. be specific, given...
Deep convolutional neural network based video super-resolution (SR) models have achieved significant progress in recent years. Existing deep SR methods usually impose optical flow to wrap the neighboring frames for temporal alignment. However, accurate estimation of is quite difficult, which tends produce artifacts super-resolved results. To address this problem, we propose a novel end-to-end that dynamically generates spatially adaptive filters alignment, are constituted by local...
As a recent noticeable topic, domain generalization (DG) aims to first learn generic model on multiple source domains and then directly generalize an arbitrary unseen target without any additional adaption. In previous DG models, by generating virtual data supplement observed domains, the augmentation based methods have shown its effectiveness. To simulate possible most of them enrich diversity original via image-level style transformation. However, we argue that potential styles are hard be...
Few-shot learning, especially few-shot image classification, has received increasing attention and witnessed significant advances in recent years. Some studies implicitly show that many generic techniques or "tricks", such as data augmentation, pre-training, knowledge distillation, self-supervision, may greatly boost the performance of a learning method. Moreover, different works employ software platforms, backbone architectures input sizes, making fair comparisons difficult practitioners...
Domain generalization (DG) aims to learn a model that generalizes well unseen target domains utilizing multiple source without re-training. Most existing DG works are based on convolutional neural networks (CNNs). However, the local operation of convolution kernel makes focus too much representations (e.g., texture), which inherently causes more prone overfit and hampers its ability. Recently, several MLP-based methods have achieved promising results in supervised learning tasks by global...
Purpose: Stanford Type B Aortic Dissection (TBAD), a critical aortic disease, has exhibited stable mortality rates over the past decade. However, diagnostic approaches for TBAD during routine health check-ups are currently lacking. This study focused on developing model to improve diagnosis in population. Patients and Methods: Serum biomarkers were investigated 88 participants using proteomic profiling combined with machine learning. The findings validated ELISA other 80 participants....
We have witnessed remarkable progress in foundation models vision tasks. Currently, several recent works utilized the segmenting anything model (SAM) to boost segmentation performance medical images, where most of them focus on training an adaptor for fine-tuning a large amount pixel-wise annotated images following fully supervised manner. In this paper, reduce labeling cost, we investigate novel weakly-supervised SAM-based model, namely WeakMedSAM. Specifically, our proposed WeakMedSAM...
The core idea of metric-based few-shot image classification is to directly measure the relations between query images and support classes learn transferable feature embeddings. Previous work mainly focuses on image-level representations, which actually cannot effectively estimate a class's distribution due scarcity samples. Some recent shows that local descriptor based representations can achieve richer than representations. However, such works are still less effective instance-level metric,...
Person retrieval faces many challenges including cluttered background, appearance variations (e.g., illumination, pose, occlusion) among different camera views and the similarity person's images. To address these issues, we put forward a novel mask based deep ranking neural network with skipped fusing layer. Firstly, to alleviate problem of masked images only foreground regions are incorporated as input in proposed network. Secondly, reduce impact variations, multi-layer fusion scheme is...
In this paper, we focus on the semi-supervised person re-identification (Re-ID) case, which only has intra-camera (within-camera) labels but not inter-camera (cross-camera) labels. real-world applications, these can be readily captured by tracking algorithms or few manual annotations, when compared with cross-camera it is very difficult to explore relationships between persons in training stage due lack of label information. To deal issue, propose a novel Progressive Cross-camera Soft-label...
As the population becomes older worldwide, accurate computer-aided diagnosis for Alzheimer's disease (AD) in early stage has been regarded as a crucial step neurodegeneration care recent years. Since it extracts low-level features from neuroimaging data, previous methods this classification problem that ignored latent featurewise relation. However, is known multiple brain regions human are anatomically and functionally interlinked according to current neuroscience perspective. Thus,...
Most existing style transfer methods follow the assumption that styles can be represented with global statistics (e.g., Gram matrices or covariance matrices), and thus address problem by forcing output images to have similar statistics. An alternative is of local patterns, where algorithms are designed swap features content images. However, limitation these they neglect semantic structure image which may lead corrupted in output. In this paper, we make a new from same region form manifold an...