SCShenghua Cheng

Shenghua Cheng

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Shenghua Cheng has a total of 119 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • CytoGPT: A multimodal generative assistant for interpretable cervical cytology analysis

    by Shijie Liu, Hongtao Kang, Hui Qiao, Yuwei Xiao, Qi Hong, Qiqi Lu, Junbo Hu, He Wu, Hua Ye, Ziheng Yu, et al.

  • Open Access

    A foundation model for multi-task cross-distribution restoration of fluorescence microscopy images

    by Qiqi Lu, Xiuli Liu, Qianjin Feng, Shaoqun Zeng, Shenghua Cheng

    Deep learning has demonstrated remarkable abilities in restoring fluorescence microscopy images degraded by noise, blur, or undersampling. However, most existing models are task-specific and trained on limited, homogeneously distributed data, which restricts their generalizability and practicality. Here, we present FluoResFM, a foundation model for multi-task and cross-distribution fluorescence microscopy image restoration in a unified model. FluoResFM leverages textual prior information to adapt to specific tasks and data distributions. Trained on datasets across three tasks (image denoising, deconvolution, and super-resolution) and over 20 biological structures, FluoResFM demonstrates superior restoration performance and enhanced generalization across datasets with varied biological structures and imaging conditions. Through fine-tuning with only a single sample, FluoResFM can further improve its performance on unseen data, achieving results comparable to conventional models trained on hundreds of samples, and be easily adapted to additional tasks, including 3D image restoration, surface projection, isotropic reconstruction, and super-resolution with various scale factors. Moreover, the performance of existing cell/organelle segmentation models can be enhanced using the high-quality images restored by FluoResFM.


  • Open Access

    The Impact of Cervical Cytology Category Imbalance on Self-Supervised Representation Learning

    by Shijie Liu, Xuben Chang, Jiaxin Bai, Li Chen, Shaoqun Zeng, Tingwei Quan, Xiuli Liucor, Shenghua Cheng

    The method of pre-training on extensive unlabeled data, followed by transferring the learned representations to downstream tasks with limited labeled data, has been effective in various fields. However, this paradigm faces the challenge of extreme data imbalance in cervical cytology, with positive cells in whole-slide images constituting approximately 1%. In this paper, we propose a pipeline for investigating the impact of this extreme category imbalance on self-supervised representation learning (SSRL). The pipeline consists of 2 stages: SSRL and downstream tasks. In the SSRL stage, we employ 2 well-established methods, masked autoencoders and the simple framework for contrastive learning, across 9 datasets with varying degrees of imbalance. The pre-trained representations are then transferred to downstream tasks by employing both linear probing and fine-tuning techniques. Additionally, we examine the effect of SSRL on annotation efficiency by varying the quantities of annotation (annotation budget). Our investigation leverages a total of 168,000 image tiles derived from 1,320 whole-slide images obtained from multiple centers. Our findings indicate a noticeable decline in accuracy (Acc) within downstream tasks as data balance shifts from 1:1 to 1:100, with a maximum drop of about 4%. This highlights the substantial impact of data imbalance on SSRL, particularly evident in downstream tasks with lower annotation rates, such as at a 1% budget. Furthermore, the downstream tasks demonstrate the potential to achieve accuracy comparable to those of scenarios with a high annotation budget (50%), even when utilizing a limited annotation budget (5%). The code is available at https://github.com/LGBluesky/ICISSRL .

  • Open Access

    CervixFM: A General Foundation Model for Cervical Cytology Image Analysis

    by Shenghua Cheng, Hua Ye, Shijie Liu, Lanjie Pei, Jiaxin Bai, Qiqi Lu, Junbo Hu, Li Chen, Jing Cai, Xi Feng, et al.

    Cervical cancer remains a significant threat to women's health globally, especially in low- and middle-income countries where cytology serves as the primary screening modality. Conventional cytology screening relies on pathologists reviewing slides, which is time-consuming and limited by inter-observer consistency. While AI-driven deep learning has advanced automated cervical cytology analysis, existing methods typically depend on abundant annotations and lack general-purpose capabilities across tasks and datasets. Faced with such challenges, we introduce CervixFM, the first general foundation model for cervical cytology image analysis. We curated a massive pretraining dataset using 18,375 whole-slide images collected from 38 batches across four medical centers, resulting in over 64 million tiles. Leveraging the large-scale data, CervixFM learns feature primitives of cervical cells via a self-supervised framework. Our model, a large vision transformer, has been systematically evaluated on five types of common analysis tasks, including slide-level classification, patch-level classification, retrieval, segmentation, and detection, and has achieved improved performance surpassing current state-of-the-art models in 34 cervical abnormality analysis. Our evaluation covers nearly all publicly available datasets in the field and involves comparisons with pretrained encoders from natural image domain and histopathology foundation models, ensuring the comprehensiveness of evaluation results. Furthermore, we investigated the label efficiency of CervixFM under low-resource supervision and assessed its cross-domain retrieval capability on six non-cervical cytology datasets. Experimental results demonstrate that CervixFM effectively captures robust representations of cervical cells, offering a new paradigm to mitigate the long-standing bottlenecks of high annotation costs and poor cross-task and cross-dataset generalizability.

  • Open Access

    A novel weakly supervised immunohistochemical cell segmentation method via counting labels

    by Hua Ye, Ziwen Zhao, Xiuli Liu, Jiabo Ma, Sibo Liu, Shaoqun Zeng, Shenghua Cheng


  • Open Access

    Pre-treatment multimodal artificial intelligence for prognostic stratification in locally advanced rectal cancer

    by Peiqi Zhuang, Shenghua Cheng, Yandong Zhao, Hongtao Kang, Chongbao Sun, Jiao Li, Wenjun Zheng, Lan Zhang, Xinhua Wang, Jing Dai, et al.

    Accurate prognostic assessment before neoadjuvant chemoradiotherapy remains challenging for locally advanced rectal cancer (LARC), limiting personalised treatment decisions. Here, we develop the Integrated Multimodal Prognostic Assessment for Locally Advanced Rectal Cancer Neoadjuvant Chemoradiotherapy (IMPACT), an artificial intelligence framework employing bidirectional multimodal attention mechanisms to capture cross-modal feature interactions and integrating pre-treatment pelvic magnetic resonance imaging, pathological biopsy whole slide images, and clinical information from 752 LARC patients across two independent centres. IMPACT achieves C-indexes of 0.805 for overall survival and 0.760 for disease-free survival, significantly outperforming the Guideline-based Imaging Risk Score (0.712 and 0.697, respectively). High-risk patients demonstrate 8.3-fold increased mortality risk and 6.5-fold increased recurrence risk compared to low-risk patients. External validation maintains robust performance with preserved risk stratification capability. Systematic ablation studies confirm the incremental value of trimodal fusion over single-modality approaches. IMPACT enables accurate pre-treatment prognostic stratification, facilitating evidence-based treatment intensification for high-risk patients and de-escalation strategies for low-risk cases in clinical practice.


  • Open Access

    A foundation model for multi-task cross-distribution restoration of fluorescence microscopy image

    by Shenghua Cheng, Qiqi Lu, Xiuli Liu, Qianjin Feng, Shaoqun Zeng

    Deep learning-based methods have demonstrated remarkable abilities in restoring high-quality fluorescence microscopy images from those degraded by noise, blur, or undersampling. However, most existing deep networks are task-specific and trained on limited, homogeneously distributed data, which restricts their generalizability and practicality in biological research. Here, we present FluoResFM, a foundation model designed for multi-task and cross-distribution fluorescence microscopy image restoration in a unified model. FluoResFM leverages textual prior information (i.e., task type, imaging object, and imaging condition) to adapt the model to specific task and data distribution. It was trained using more than 4.3 million paired patches across three tasks (image denoising, deconvolution, and super-resolution) and over 20 biological structure types. FluoResFM demonstrates superior restoration performance and enhanced generalization, delivering high-fidelity reconstruction results across the three tasks and diverse internal and unseen external datasets encompassing varied biological structures and imaging conditions. Leveraging its strong generalization capability, FluoResFM can further improve its performance on unseen data through fine-tuning with only a single sample, achieving results comparable to those of conventional deep networks trained on hundreds of samples. Furthermore, the performance of existing cell/organelle segmentation models can be further improved using the high-quality image restored by FluoResFM. To make FluoResFM widely accessible to the biology research community, we developed a user-friendly napari plugin. These establish FluoResFM as a versatile foundation model for fluorescence microscopy image processing and analysis.

  • Open Access

    A multimodal deep learning framework with contrastive learning and multi-instance learning for endometrial cancer preoperative risk stratification

    by Shenghua Cheng, Hongtao Kang, Yu Wang, Haoqiang He, Gengxi Cai, Yujian Zou, Jianpeng Li, Yi Shen, Jiayi Zhuo, Abuduxukuer Aili, et al.

    Endometrial cancer (EC) risk stratification currently relies on postoperative pathology, limiting preoperative surgical planning. Magnetic resonance imaging (MRI) is the preferred modality for preoperative EC evaluation. However, current research on EC preoperative risk stratification has certain limitations: traditional radiomics demonstrate deficient performance while relying on manual segmentation, and unimodal designs that fail to fully leverage clinical data. To address this, we developed an automated deep learning framework integrating multiparametric MRI and MRI reports from 2,662 patients across six centers. Key innovations include: contrastive learning leverages MRI reports to supervise and refine the image encoder, aligning visual features with diagnostic semantics; and multi-instance learning dynamically aggregates features from multi-parametric MRI sequences, even in the presence of missing data. Our multimodal model significantly outperformed unimodal approaches, achieving AUCs of 0.827 (internal validation) and 0.768-0.863 (four external cohorts), improving AUC by 4.7%-10.2% versus image-only and 4.1%-8.5% versus text-only models. This demonstrates strong generalizability and offers potential to optimize surgical planning, improve prognosis, and reduce complications.

  • Open Access

    AI-assisted cervical cytology precancerous screening for high-risk population in resource-limited regions using a compact microscope

    by Jiaxin Bai, Ning Li, Hua Ye, Xu Li, Li Chen, Junbo Hu, Baochuan Pang, Xiaodong Chen, Gong Rao, Qinglei Hu, et al.

    Insufficient coverage of cervical cytology screening in resource-limited areas remains a major bottleneck for women’s health, as traditional centralized methods require significant investment and many qualified pathologists. Using consumer-grade electronic hardware and aspherical lenses, we design an ultra-low-cost and compact microscope. Given the microscope’s low resolution, which hinders accurate identification of lesion cells in cervical samples, we train a coarse instance classifier to screen and extract feature sequences of the top 200 instances containing potential lesions from a slide. We further develop Att-Transformer to focus on and integrate the sparse lesion information from these sequences, enabling slide grading. Our model is trained and validated using 3510 low-resolution slides from female patients at four hospitals, and subsequently evaluated on four independent datasets. The system achieves area under the receiver operating characteristic curve values of 0.87 and 0.89 for detecting squamous intraepithelial lesions on 364 slides from female patients at two external primary hospitals, 0.89 on 391 newly collected slides from female patients at the original four hospitals, and 0.85 on 570 human papillomavirus positive slides from female patients. These findings demonstrate the feasibility of our AI-assisted approach for effective detection of high-risk cervical precancer among women in resource-limited regions.


  • Open Access

    Evaluating the performance of large language & visual-language models in cervical cytology screening

    by Qi Hong, Shijie Liu, Liying Wu, Qiqi Lu, Pinglan Yang, Dingyu Chen, Gong Rao, Xinyi Liu, Hua Ye, Peiqi Zhuang, et al.

    Large language models (LLMs) and large visual-language models (LVLMs) have exhibited near-human levels of knowledge, image comprehension, and reasoning abilities, and their performance has undergone evaluation in some healthcare domains. However, a systematic evaluation of their capabilities in cervical cytology screening has yet to be conducted. Here, we constructed CCBench, a benchmark dataset dedicated to the evaluation of LLMs and LVLMs in cervical cytology screening, and developed a GPT-based semi-automatic evaluation pipeline to assess the performance of six LLMs (GPT-4, Bard, Claude-2.0, LLaMa-2, Qwen-Max, and ERNIE-Bot-4.0) and five LVLMs (GPT-4V, Gemini, LLaVA, Qwen-VL, and ViLT) on this dataset. CCBench comprises 773 question-answer (QA) pairs and 420 visual-question-answer (VQA) triplets, making it the first dataset in cervical cytology to include both QA and VQA data. We found that LLMs and LVLMs demonstrate promising accuracy and specialization in cervical cytology screening. GPT-4 achieved the best performance on the QA dataset, with an accuracy of 70.5% for close-ended questions and average expert evaluation score of 6.9/10 for open-ended questions. On the VQA dataset, Gemini achieved the highest accuracy for close-ended questions at 67.8%, while GPT-4V attained the highest expert evaluation score of 6.1/10 for open-ended questions. Besides, LLMs and LVLMs revealed varying abilities in answering questions across different topics and difficulty levels. However, their performance remains inferior to the expertise exhibited by cytopathology professionals, and the risk of generating misinformation could lead to potential harm. Therefore, substantial improvements are required before these models can be reliably deployed in clinical practice.