Yushan Zheng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yushan Zheng has a total of 92 co-authors. Here are the top co-authors by number of shared publications:
- Jun Shi
Hefei University of Technology
23 shared publications Zhiguo Jiang
Beihang University
22 shared publications- Zhiguo Jiang
Beihang University
17 shared publications - Haopeng Zhang
Beihang University
13 shared publications Jun Shi
School of Software, Hefei University of Technology, Hefei, China
13 shared publicationsFengying Xie
Image Processing Center, School of Astronautics, Beihang University, Beijing, 100191, China; Tianmushan Laboratory, Hangzhou, 311115, Zhejiang Province, China
12 shared publications
Latest Publications
- 8 November 2026
- Published by Springer Nature
- p. 376-388
Weakly Supervised Vision-Language Learning for EGFR Mutation Status Prediction of Lung Cancer from CT Image and Report
- 8 November 2026
- Published by Springer Nature
- p. 376-388
Weakly Supervised Vision-Language Learning for EGFR Mutation Status Prediction of Lung Cancer from CT Image and Report
- 10 July 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 113, 104213
Adapting pathology foundation models for continual cross-center WSI retrieval
- 10 July 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 113, 104213
Adapting pathology foundation models for continual cross-center WSI retrieval
- 18 March 2026
Token-guided multimodal prognosis in hepatocellular carcinoma: a framework steered by tumour-stroma ratio.
The tumour-stroma ratio (TSR) is a potential prognostic indicator, yet hindered by quantification challenges and conflicting reports. To determine whether TSR follows a non-linear prognostic pattern and to develop an artificial intelligence (AI)-powered framework for standardised TSR assessment and prognosis prediction in hepatocellular carcinoma (HCC). We integrated whole-slide image (WSI) data with clinical variables across a retrospective cohort (n=392) and The Cancer Genome Atlas dataset (n=168). Restricted cubic splines were used to interrogate non-linear hazard dynamics, with biological validation via transcriptomics and immunohistochemistry. An AI-driven foundation model framework was developed for TSR quantification and multimodal prognostic modelling. Our analysis unveiled an inverted U-shaped non-linear relationship between TSR and mortality, identifying a risk initiation threshold at 0.188 and a peak at 0.268. Transcriptomics analysis indicated that this high-risk phenotype is characterised by active tumour proliferation, stromal activation and tumour microenvironment crosstalk. Technically, AI-derived TSR showed strong correlation with expert assessment (R² >0.9). Furthermore, we developed a novel 'Token-Guided Multimodal Fusion' architecture to integrate WSI, TSR and clinical variables as high-dimensional tokens directly into the computational logic. Consequently, our multimodal framework demonstrated prognostic accuracy (area under the curve >0.80) compared with unimodal baselines. This study redefines TSR assessment, shifting from manual estimation to high-dimensional semantic reasoning. By identifying the non-linear prognostic mechanics of the stroma, our token-guided framework offers a biologically interpretable solution for HCC. We suggest that the future of computational pathology may lie not in simple quantification, but in the semantic fusion of human domain knowledge with AI reasoning.
- 18 March 2026
Token-guided multimodal prognosis in hepatocellular carcinoma: a framework steered by tumour-stroma ratio.
The tumour-stroma ratio (TSR) is a potential prognostic indicator, yet hindered by quantification challenges and conflicting reports. To determine whether TSR follows a non-linear prognostic pattern and to develop an artificial intelligence (AI)-powered framework for standardised TSR assessment and prognosis prediction in hepatocellular carcinoma (HCC). We integrated whole-slide image (WSI) data with clinical variables across a retrospective cohort (n=392) and The Cancer Genome Atlas dataset (n=168). Restricted cubic splines were used to interrogate non-linear hazard dynamics, with biological validation via transcriptomics and immunohistochemistry. An AI-driven foundation model framework was developed for TSR quantification and multimodal prognostic modelling. Our analysis unveiled an inverted U-shaped non-linear relationship between TSR and mortality, identifying a risk initiation threshold at 0.188 and a peak at 0.268. Transcriptomics analysis indicated that this high-risk phenotype is characterised by active tumour proliferation, stromal activation and tumour microenvironment crosstalk. Technically, AI-derived TSR showed strong correlation with expert assessment (R² >0.9). Furthermore, we developed a novel 'Token-Guided Multimodal Fusion' architecture to integrate WSI, TSR and clinical variables as high-dimensional tokens directly into the computational logic. Consequently, our multimodal framework demonstrated prognostic accuracy (area under the curve >0.80) compared with unimodal baselines. This study redefines TSR assessment, shifting from manual estimation to high-dimensional semantic reasoning. By identifying the non-linear prognostic mechanics of the stroma, our token-guided framework offers a biologically interpretable solution for HCC. We suggest that the future of computational pathology may lie not in simple quantification, but in the semantic fusion of human domain knowledge with AI reasoning.
- 22 January 2026
- Published by Elsevier in Pattern Recognition
- Vol. 175
Lifelong content-based histopathology image retrieval via bilevel coreset selection and distance consistency rehearsal
- 22 January 2026
- Published by Elsevier in Pattern Recognition
- Vol. 175
Lifelong content-based histopathology image retrieval via bilevel coreset selection and distance consistency rehearsal
- Open Access
EndoCIL: A Class-Incremental Learning Framework for Endoscopic Image Classification
Class-incremental learning (CIL) for endoscopic image analysis is crucial for real-world clinical applications, where diagnostic models should continuously adapt to evolving clinical data while retaining performance on previously learned ones. However, existing replay-based CIL methods fail to effectively mitigate catastrophic forgetting due to severe domain discrepancies and class imbalance inherent in endoscopic imaging. To tackle these challenges, we propose EndoCIL, a novel and unified CIL framework specifically tailored for endoscopic image diagnosis. EndoCIL incorporates three key components: Maximum Mean Discrepancy Based Replay (MDBR), employing a distribution-aligned greedy strategy to select diverse and representative exemplars, Prior Regularized Class Balanced Loss (PRCBL), designed to alleviate both inter-phase and intra-phase class imbalance by integrating prior class distributions and balance weights into the loss function, and Calibration of Fully-Connected Gradients (CFG), which adjusts the classifier gradients to mitigate bias toward new classes. Extensive experiments conducted on four public endoscopic datasets demonstrate that EndoCIL generally outperforms state-of-the-art CIL methods across varying buffer sizes and evaluation metrics. The proposed framework effectively balances stability and plasticity in lifelong endoscopic diagnosis, showing promising potential for clinical scalability and deployment.
Open AccessEndoCIL: A Class-Incremental Learning Framework for Endoscopic Image Classification
Class-incremental learning (CIL) for endoscopic image analysis is crucial for real-world clinical applications, where diagnostic models should continuously adapt to evolving clinical data while retaining performance on previously learned ones. However, existing replay-based CIL methods fail to effectively mitigate catastrophic forgetting due to severe domain discrepancies and class imbalance inherent in endoscopic imaging. To tackle these challenges, we propose EndoCIL, a novel and unified CIL framework specifically tailored for endoscopic image diagnosis. EndoCIL incorporates three key components: Maximum Mean Discrepancy Based Replay (MDBR), employing a distribution-aligned greedy strategy to select diverse and representative exemplars, Prior Regularized Class Balanced Loss (PRCBL), designed to alleviate both inter-phase and intra-phase class imbalance by integrating prior class distributions and balance weights into the loss function, and Calibration of Fully-Connected Gradients (CFG), which adjusts the classifier gradients to mitigate bias toward new classes. Extensive experiments conducted on four public endoscopic datasets demonstrate that EndoCIL generally outperforms state-of-the-art CIL methods across varying buffer sizes and evaluation metrics. The proposed framework effectively balances stability and plasticity in lifelong endoscopic diagnosis, showing promising potential for clinical scalability and deployment.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (11) , 4267-4280
Slide-Based Graph Collaborative Training for Histopathology Whole Slide Image Analysis
The development of computational pathology lies in the consensus that pathological characteristics of tumors are significant guidance for cancer diagnostics. Most existing research focuses on the inner-contextual information within each WSI yet ignores the possible inter-correlations between slides. As the development of tumors is a continuous process involving a series of histological, morphological, and genetic changes that accumulate over time, the similarities and differences between WSIs across various stages, grades, locations and patients should potentially contribute to the representation of WSIs and deserve to be taken into account in WSI modeling. To verify the advancement of introducing the slide inter-correlations into the representation learning of WSIs, we proposed a generic WSI analysis pipeline SlideGCD that can be adapted to any existing Multiple Instance Learning (MIL) frameworks and improve their performance. With the new paradigm, the prior knowledge of cancer development can participate in the end-to-end workflow, which concurrently initializes and refines the slide representation, as a guide for message passing in the slide-based graph. Extensive comparisons and experiments are conducted to validate the effectiveness and robustness of the proposed pipeline across 4 different tasks, including cancer subtyping, cancer staging, survival prediction, and genemutation prediction, with 8 representative SOTA WSI analysis frameworks as backbones. The code is available at https://github.com/HFUT-miaLab/SlideGCD.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (11) , 4267-4280
Slide-Based Graph Collaborative Training for Histopathology Whole Slide Image Analysis
The development of computational pathology lies in the consensus that pathological characteristics of tumors are significant guidance for cancer diagnostics. Most existing research focuses on the inner-contextual information within each WSI yet ignores the possible inter-correlations between slides. As the development of tumors is a continuous process involving a series of histological, morphological, and genetic changes that accumulate over time, the similarities and differences between WSIs across various stages, grades, locations and patients should potentially contribute to the representation of WSIs and deserve to be taken into account in WSI modeling. To verify the advancement of introducing the slide inter-correlations into the representation learning of WSIs, we proposed a generic WSI analysis pipeline SlideGCD that can be adapted to any existing Multiple Instance Learning (MIL) frameworks and improve their performance. With the new paradigm, the prior knowledge of cancer development can participate in the end-to-end workflow, which concurrently initializes and refines the slide representation, as a guide for message passing in the slide-based graph. Extensive comparisons and experiments are conducted to validate the effectiveness and robustness of the proposed pipeline across 4 different tasks, including cancer subtyping, cancer staging, survival prediction, and genemutation prediction, with 8 representative SOTA WSI analysis frameworks as backbones. The code is available at https://github.com/HFUT-miaLab/SlideGCD.
- Open Access
- 18 April 2025
- Published by Springer Nature in Breast Cancer Research
- Vol. 27 (1) , 1-10
Machine learning prediction of HER2-low expression in breast cancers based on hematoxylin–eosin-stained slides
Treatment with HER2-targeted therapies is recommended for HER2-positive breast cancer patients with HER2 gene amplification or protein overexpression. Interestingly, recent clinical trials of novel HER2-targeted therapies demonstrated promising efficacy in HER2-low breast cancers, raising the prospect of including a HER2-low category (immunohistochemistry, IHC) score of 1 + or 2 + with non-amplified in-situ hybridization for HER2-targeted treatments, which necessitated the accurate detection and evaluation of HER2 expression in tumors. Traditionally, HER2 protein levels are routinely assessed by IHC in clinical practice, which not only requires significant time consumption and financial investment but is also technically challenging for many basic hospitals in developing countries. Therefore, directly predicting HER2 expression by hematoxylin-eosin (HE) staining should be of significant clinical values, and machine learning may be a potent technology to achieve this goal. In this study, we developed an artificial intelligence (AI) classification model using whole slide image of HE-stained slides to automatically assess HER2 status. A publicly available TCGA-BRCA dataset and an in-house USTC-BC dataset were applied to evaluate our AI model and the state-of-the-art method SlideGraph + in terms of accuracy (ACC), the area under the receiver operating characteristic curve (AUC), and F1 score. Overall, our AI model achieved the superior performance in HER2 scoring in both datasets with AUC of 0.795 ± 0.028 and 0.688 ± 0.008 on the USCT-BC and TCGA-BRCA datasets, respectively. In addition, we visualized the results generated from our AI model by attention heatmaps, which proved that our AI model had strong interpretability. Our AI model is able to directly predict HER2 expression through HE images with strong interpretability, and has a better ACC particularly in HER2-low breast cancers, which provides a method for AI evaluation of HER2 status and helps to perform HER2 evaluation economically and efficiently. It has the potential to assist pathologists to improve diagnosis and assess biomarkers for companion diagnostics.
- 18 April 2025
- Published by Springer Nature in Breast Cancer Research
- Vol. 27 (1) , 1-10
Open AccessMachine learning prediction of HER2-low expression in breast cancers based on hematoxylin–eosin-stained slides
Treatment with HER2-targeted therapies is recommended for HER2-positive breast cancer patients with HER2 gene amplification or protein overexpression. Interestingly, recent clinical trials of novel HER2-targeted therapies demonstrated promising efficacy in HER2-low breast cancers, raising the prospect of including a HER2-low category (immunohistochemistry, IHC) score of 1 + or 2 + with non-amplified in-situ hybridization for HER2-targeted treatments, which necessitated the accurate detection and evaluation of HER2 expression in tumors. Traditionally, HER2 protein levels are routinely assessed by IHC in clinical practice, which not only requires significant time consumption and financial investment but is also technically challenging for many basic hospitals in developing countries. Therefore, directly predicting HER2 expression by hematoxylin-eosin (HE) staining should be of significant clinical values, and machine learning may be a potent technology to achieve this goal. In this study, we developed an artificial intelligence (AI) classification model using whole slide image of HE-stained slides to automatically assess HER2 status. A publicly available TCGA-BRCA dataset and an in-house USTC-BC dataset were applied to evaluate our AI model and the state-of-the-art method SlideGraph + in terms of accuracy (ACC), the area under the receiver operating characteristic curve (AUC), and F1 score. Overall, our AI model achieved the superior performance in HER2 scoring in both datasets with AUC of 0.795 ± 0.028 and 0.688 ± 0.008 on the USCT-BC and TCGA-BRCA datasets, respectively. In addition, we visualized the results generated from our AI model by attention heatmaps, which proved that our AI model had strong interpretability. Our AI model is able to directly predict HER2 expression through HE images with strong interpretability, and has a better ACC particularly in HER2-low breast cancers, which provides a method for AI evaluation of HER2 status and helps to perform HER2 evaluation economically and efficiently. It has the potential to assist pathologists to improve diagnosis and assess biomarkers for companion diagnostics.
- Open Access
- 11 April 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 39 (7) , 7247-7256
Promptable Representation Distribution Learning and Data Augmentation for Gigapixel Histopathology WSI Analysis
Gigapixel image analysis, particularly for whole slide images (WSIs), often relies on multiple instance learning (MIL). Under the paradigm of MIL, patch image representations are extracted and then fixed during the training of the MIL classifiers for efficiency consideration. However, the invariance of representations makes it difficult to perform data augmentation for WSI-level model training, which significantly limits the performance of the downstream WSI analysis. The current data augmentation methods for gigapixel images either introduce additional computational costs or result in a loss of semantic information, which is hard to meet the requirements for efficiency and stability needed for WSI model training. In this paper, we propose a Promptable Representation Distribution Learning framework (PRDL) for both patch-level representation learning and WSI-level data augmentation. Meanwhile, we explore the use of prompts to guide data augmentation in feature space, which achieves promptable data augmentation for training robust WSI-level models. The experimental results have demonstrated that the proposed method stably outperforms state-of-the-art methods.
- 11 April 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 39 (7) , 7247-7256
Open AccessPromptable Representation Distribution Learning and Data Augmentation for Gigapixel Histopathology WSI Analysis
Gigapixel image analysis, particularly for whole slide images (WSIs), often relies on multiple instance learning (MIL). Under the paradigm of MIL, patch image representations are extracted and then fixed during the training of the MIL classifiers for efficiency consideration. However, the invariance of representations makes it difficult to perform data augmentation for WSI-level model training, which significantly limits the performance of the downstream WSI analysis. The current data augmentation methods for gigapixel images either introduce additional computational costs or result in a loss of semantic information, which is hard to meet the requirements for efficiency and stability needed for WSI model training. In this paper, we propose a Promptable Representation Distribution Learning framework (PRDL) for both patch-level representation learning and WSI-level data augmentation. Meanwhile, we explore the use of prompts to guide data augmentation in feature space, which achieves promptable data augmentation for training robust WSI-level models. The experimental results have demonstrated that the proposed method stably outperforms state-of-the-art methods.