🎉New!!!🎉Online Self-distillation and Self-modeling for 3D Brain Tumor Segmentation

Yan Pang, Yunhao Li, Teng Huang*, Jiaming Liang, Zhen Wang, Changyu Dong, Dongyang Kuang, Ying Hu, Hao Chen, Tim Lei, Qiong Wang*

Project     Journal (JCR Q1, IF 7.7)

• MOD is a plug-and-play component designed to enhance supervised learning models for brain tumor segmentation in data-scarce environments by integrating an Online Tokenizer and a Dense Predictor. These elements facilitate efficient representation learning through self-distillation and self-modeling on masked patches, promoting swift convergence.

🎉New!!!🎉Efficient Breast Lesion Segmentation from Ultrasound Videos Across Multiple Source-limited Platforms

Yan Pang, Yunhao Li, Teng Huang*, Jiaming Liang, Ziyu Ding, Hao Chen, Baoliang Zhao, Ying Hu, Zheng Zhang, Qiong Wang*

Project     Journal (JCR Q1, IF 7.7)

• BaS is an innovative, on-device breast lesion segmentation model designed for medical videos, integrating Stem module and BaSBlock to refine representations through inter- and intra-frame analysis, thus balancing segmentation performance and inference speed. This model surpasses existing top-performing models in efficiency and prediction accuracy on resource-limited devices by offering two versions, BaS-S for enhanced segmentation accuracy and BaS-L for faster inference times.

Comprehensive Transformer Integration Network (CTIN): Advancing Endoscopic Disease Segmentation with Hybrid Transformer Architecture

Jiaming Liang, Mingdu Zhang, Caiyan Tan, Teng Huang, Xi Zhang, Zheng Zhang*, Shegan Gao, Qian Sheng*, Yan Pang*

Conference paper (CCF-C)

• CTIN, leveraging Transformer’s strengths and introducing a Global Local Fusion Extractor module, enhances the precision of endoscopic image segmentation by preserving boundary details and accurately localizing objects, thus demonstrating superior performance and robustness in challenging scenarios. This approach addresses the limitations of existing models in capturing complex boundaries and small objects, achieving state-of-the-art results on various metrics.

Lightweight Multispectral Skeleton and Multi-stream Graph Attention Networks for Enhanced Action Prediction with Multiple Modalities

Teng Huang*, Weiqing Kong, Jiaming Liang, Ziyu Ding, Hui Li, Xi Zhang*

Conference paper (CCF-C)

• LMS-GAT introduces a novel approach to human action recognition by facilitating information exchange through node concentration and diffusion across structural and temporal dimensions, selectively suppressing and reinstating structural node representations for specific actions ,and utilizing hierarchical shifted temporal windows to assess temporal information. This method effectively addresses dynamic changes during action transitions, outperforming state-of-the-art methods in prediction accuracy on the NTU RGB+D 60 and 120 datasets, thus highlighting its improved efficacy in capturing and recognizing human actions.

AdaptFormer: An Adaptive Hierarchical Semantic Approach for Change Detection on Remote Sensing Images

Teng Huang, Yile Hong, Yan Pang, Jiaming Liang, Jie Hong, Lin Huang, Yuan Zhang, Yan Jia, Patrizia Savi

Project     Journal (JCR Q1, IF 5.6)

• AdaptFormer is designed to adaptively interpret hierarchical semantics in remote sensing change detection, employing tailored strategies across three semantic depths to enhance the identification of environmental changes, from straightforward operations for shallow semantics to cascaded depthwise attention for in-depth semantics. This approach enables AdaptFormer to surpass leading benchmarks, achieving exceptional F1 and IoU scores of 92.65% and 86.31% on the LEVIR-CD dataset, and 97.59% and 95.29% on the DSIFN-CD dataset, respectively.

AgileNet: A Rapid and Efficient Breast Lesion Segmentation Method for Medical Image Analysis

Jiaming Liang, Teng Huang*, Dan Li, Ziyu Ding, Yunhao Li, Lin Huang, Qiong Wang, Xi Zhang*

Conference paper (CCF-C)

• AgileNet is an efficient lesion segmentation model that balances accuracy and efficiency by combining the strengths of convolutional neural networks and transformers, utilizing an Agile block for cost-effective information exchange that incorporates both global and local contexts. Demonstrating superior performance in terms of accuracy, model size, and throughput on resource-constrained devices, AgileNet offers a promising solution for accurate and efficient medical image segmentation in such settings.

Slim UNETR: Scale Hybrid Transformers to Efficient 3D Medical Image Segmentation Under Limited Computational Resources

Yan Pang (Adviser), Jiaming Liang, Teng Huang*, Hao Chen, Yunhao Li, Dan Li, Lin Huang, Qiong Wang*

Project     Journal (JCR Q1, IF 10.6)

• SlimUNETR is a lightweight 3D medical image segmentation model that achieves state-of-the-art performance on the BraTS2021 and other 3D lesion segmentation datasets, while also seeing a significant reduction in the number of parameters and floating-point computations.

GanNoise: Defending against black-box membership inference attacks by countering noise generation

Jiaming Liang, Teng Huang*, Zidan Luo, Dan Li, Yunhao Li, Ziyu Ding

Conference paper (EI)

• GanNoise is a novel training framework designed to preserve privacy in deep learning models by generating noise that adds randomness to private data during training, thereby preventing excessive memorization of the actual training data while maintaining classification accuracy. This approach effectively defends against membership inference attacks without compromising model performance, outperforming existing advanced MIA defense solutions in terms of efficiency across various datasets.

Cam-PC: A Novel Method for Camouflaging Point Clouds to Counter Adversarial Deception in Remote Sensing

Bo Wei, Teng Huang*, Xi Zhang, Jiaming Liang, Yunhao Li, Ziyu Ding, Cong Cao, Dan Li, Yongfeng Chen, Huagang Xiong, Feng Jiang, Xiqiu Zhang*

Journal (JCR Q2, IF 4.7)

• This article explores the increased privacy risks associated with multi-task learning (MTL) compared to traditional single-task learning, through the design of model extraction attacks (MEA) and membership inference attacks (MIA) in MTL. Evaluations across six MTL model architectures and two popular MTL datasets reveal that both the number of tasks and the complexity of training data significantly influence attack performance, demonstrating that MTL is more vulnerable to these attacks than single-task learning. This highlights the need for enhanced privacy measures in MTL frameworks.

MTL-Leak: Privacy Risk Assessment in Multi-Task Learning

Hongyang Yan, Anli Yan, Li Hu, Jiaming Liang, Haibo Hu*

Journal (JCR Q2, IF 7.0)

• Multi-task learning (MTL) supports simultaneous training over multiple related tasks and learns the shared representation. While improving the generalization ability of training on a single task, MTL has higher privacy risk than traditional single-task learning because more sensitive information is extracted and learned in a correlated manner. Unfortunately, very few works have attempted to address the privacy risks posed by MTL. In this article, we first investigate such risk by designing model extraction attack (MEA) and membership inference attack (MIA) in MTL. Then we evaluate the privacy risks on six MTL model architectures and two popular MTL datasets, whose results show that both the number of tasks and the complexity of training data play an important role in the attack performance. Our investigation shows that MTL is more vulnerable than traditional single-task learning under both attacks.