Computational Digital Pathology Lab

The global shortage of pathologists and rising cancer incidences have strained healthcare resources, leading to delayed and inadequate patient care. Our AI-based digital pathology (AIDP) Programme addresses these challenges by providing fast, accurate, and accessible diagnostic tools. This prorgamme aims to overcome ecosystem data and infrastructure issues, lack of production pipelines, smalle and low quality database and poor scalability of AI diagnostic models across different sites. By establishing an AI-powered diagnostic solutions and developing necessary intellectual properties like A!HistoClouds and A!MagQC as demonstrated in Figure 1, we support AI model optimization and generalization, such as A!Prostate model. Additionally, creating a large-scale annotated DP database and standardized annotation schemes enhances the development and deployment of AI solutions in clinical settings.

Figure 1. AI-based Digital Pathology Diagnosis workflow and its components, showcasing the innovative approaches for improving diagnostic accuracy and efficiency in pathology. A) The A!MagQC highlights the transition from manual assessment of pathology slides, which is often tedious and subjective, to an automated approach. It emphasizes the importance of high-quality data in ensuring the accuracy of AI models for cytology and pathology. The panel shows how AI can identify quality issues affecting interpretations and diagnoses, making the process more efficient and accurate. B). The A!HistoClouds  demonstrates a cloud-based annotation solution for pathology. It supports various types of annotations, including normal tissue, benign disease, and malignant regions. The panel highlights the integration of multiple tools for streamlined image annotation, showcasing a significant increase in AI-generated annotations. This module facilitates collaborative efforts and enhances the efficiency of the annotation process through cloud-based technology. C). The A!Prostate model presents a cloud-based annotation solution specifically designed for prostate Gleason Grading. It employs convolutional neural networks (CNNs) to distinguish between different grades of prostate cancer, including stroma, benign, Gleason 3 - 5. The panel illustrates how this AI-driven solution provides detailed and accurate annotations, contributing to better diagnostic precision and improved patient outcomes.

Our AIDP program’s goals include Image Appearance Migration (IAM) for cytology/pathology samples, developing a framework for AI models such as A!Prostate for Gleason grading, and utilizing generative AI to reduce dependence on human annotations. We also plan to develop autonomous cycled reinforcement learning to improve AI model explainability and generalization. By validating clinical AI pathological diagnostic models through international deployment and local cross-validation, we ensure high performance and accuracy. Our collaboration with institutions like A*STAR, NUH, SGH, and TTSH allows us to develop a comprehensive digital pathology development and adoption platform, integrating AI with a real-world pathological database. This initiative promises to enhance healthcare efficiency, generate commercially appealing intellectual properties, and stimulate job creation and talent development, positioning Singapore as a leading hub for AI-based diagnostics.

Establishing a Comprehensive Pipeline for Medical-Class Digital Pathology Diagnostic Models

Figure 4. A!HistoClouds emphasizes its role in fighting cancer by leveraging AI models and pathology images for enhanced diagnostics and analysis. The central image shows the A!HistoClouds interface, accessible on various devices, including smartphones and desktop computers. This interface allows users to view proprietary bright field and fluorescence virtual slide formats, perform annotations, and analyze results directly in their browser from any device. The top right corner depicts a researcher and a robotic assistant using A!HistoClouds to annotate and analyze pathology images. The cloud infrastructure in the middle signifies the cloud-based nature of the solution, enabling seamless access to data and computational resources. The bottom right corner shows a medical professional reviewing annotated prostate slides, highlighting the collaboration between AI and human expertise. The AI annotations are color-coded to indicate different tissue types and diagnostic features, enhancing the accuracy and efficiency of cancer diagnosis.

Enhancing AI Diagnostic Accuracy Through Adaptive Generalization Techniques – A!Prostate Model

Figure 5. The comprehensive overview of the AI-based Digital Pathology Diagnosis system’s performance and validation across various stages and conditions. A). Shows the input of original pathology images from various acquisition systems and the consistent output results generated by the our AI model, ensuring reliable performance across different imaging platforms. B). Displays the transformation of the original images into consistent AI-generated results, demonstrating the robustness of the AI model in handling diverse input sources. C). Presents a bar graph comparing the macro average F1 scores of different scanners, with and without IAM generalization techniques. The results indicate that generalization techniques significantly enhance the model’s performance across different scanners. D). Outlines the three phases of the validation process: Phase 1: Traditional microscopic examination; Phase 2: Whole slide image (WSI) examination without AI assistance; Phase 3: WSI examination with AI assistance, highlighting the improvement in diagnostic efficiency and accuracy with AI integration. E). Shows a bar graph of the quadratic weighted kappa scores across different pathologists and phases, illustrating the consistency and reliability of AI-assisted diagnostics. F). Depicts a box plot comparing the examination time per WSI for each pathologist across the three phases, showing a significant reduction in examination time with AI assistance. G). Illustrates a specific case of cancer screening and Gleason grading using the AI model. The AI’s output is compared with the ground truth (GT), showing a 100% true positive rate for cancer screening and a high quadratic weighted kappa score of 0.845 for Gleason grading. This panel also emphasizes the ongoing large-scale, multi-center validation efforts to further confirm the AI model’s effectiveness.

Development and training of our AI-enhanced solution began when scanned images were uploaded to A!HistoClouds, where clinical pathologists annotated regions of interest (ROIs) and assigned labels to them (stroma, normal glands, Gleason Pattern 3-5, for example in prostate diagnosis). After cropping the images into small patches and splitting the data into training and testing sets, we trained a convolutional neural network (CNN) using this training dataset to classify image patches into different label categories. On a per-image patch level, the sensitivity, specificity, positive predictive value, and negative predictive value were 93.0%, 92.7%, 96.3%, and 81.5%, respectively. We then visualized the prediction results on whole slide level by applying the model to the entire image. Comparing the generated heat maps and ground truth (pathologists’ annotations and Gleason scores), the predictions made by the A!Prostate model (https://www.nature.com/articles/s43856-024-00502-1) were well-matched with the pathologists’ diagnoses. For every different scanner brand, the digitization techniques and parameters are different. Therefore, the images of a single glass slide scanned by different scanners have varied appearances. To overcome this variation issue, we developed an adaptive solution by applying generalization techniques, named as Image Appearance Migration (IAM) to generate consistent results. Our strategy was to standardize images across these scanners and reduce image inconsistencies. Additionally, we further enhanced the generalizability of the models by enlarging the original training dataset using colour augmentation to simulate the appearance variations of different scanners during training. After applying generalization techniques, our AI models demonstrated increased accuracy and consistency among different scanners, potentially other variation in the upstream steps such as staining. Our experimental results indicate that our model is independent of the scanner hardware used, demonstrating its generalization ability across different scanning devices. For a video introduction, please refer to the YouTube: https://youtu.be/M8gPByfyQMU?si=19RwONOCYLVlSoAP  

• Xinmi Huo^#, Kok Haur Ong^#, Kah Weng Lau, Laurent Gole, Char Loo Tan, Chongchong Zhang, Yonghui Zhang, Xiaohui Zhu, Longjie Li, Hao Han, David Young, Haoda Lu, Jun Xu, Wanyuan Chen, Stephan J Sanders, Lee Hwee Kuan, Susan Swee-Shan Hue^&, Weimiao YU^&, Soo Yong Tan^&. Comprehensive AI Model Development for Gleason Grading: From Scanning, Cloud-Based Annotation to Pathologist-AI Interaction. (2024). Nature - Communication Medicine. (IF NA)


• Chunyue FENG^#, Kokhaur ONG^#, David M. YOUNG^#, Bingxian CHEN, Longjie LI, Weizhong GU, Fei LIU, Hongfeng TANG, Manli ZHAO, Min YANG, Kun ZHU, Limin HUANG, Qiang WANG, Xinmi HUO, Gabriel P.L. MARINI, Haoda LU, Kun GUI, Hao HAN, Stephan J. SANDERS, Lin LI^&, Weimiao YU^&, and Jianhua MAO^&. Artificial intelligence-assisted quantification and assessment of whole slide images for paediatric kidney disease diagnosis. (2024). Bioinformatics. (IF = 6.9)


• Emanuela Frittoli^#, Andrea Palamidessi^#, Fabio Iannelli, Federica Zanardi, Stefano Villa, Leonardo Barzaghi, Hind Ando, Valeria Cancila, Galiba Beznuskenko, Giulia Della Chiara, Massimiliano Pagani, Chiara Malinverno, Dipanjan Bhattacharya, Federica Pisati, Weimiao Yu, Viviana Galimberti, Giuseppina Bonizzi, Emanuele Martini, Alexander Mironov, Chiara Rossi, Giovanni Bertalot, Marco Lucioni, Cristiuano Perini, FRancesco Ferrari, Chiara Lanzuolo, Guilherme Nader, Marco Foiani, Matthieu Piel, Roberto Cerbino, Fabio Giavazzi^&, Claudio Tripodo^&, Giorgio Scita^&. Tissue fluidification promotes a cGAS/STING-mediated cytosolic DNA response in invasive breast cancer. (2023). Nature Material. (IF = 41.2)


• Longjie Li^#, Arun Mouli Kolinjivadi^#, Kok Haur Ong, David M Young, Gabriel Pik Liang Marini, Sock Hoai Chan, Siao Ting Chong, Ee Ling Chew, Haoda Lu, Laurent Gole^&, Weimiao Yu^&, Joanne Ngeow^&. Automatic DNA replication tract measurement to assess replication and repair dynamics at the single-molecule level. (2022). Bioinformatics. (IF = 6.9)


• Xiaohui Zhu^#, Xiaoming Li^#, Kokhaur Ong^#, Wenli Zhang, Wencai Li, Longjie Li, David Young, Yongjian Su, Bin Shang, Linggan Peng, Wei Xiong, Yunke Liu, Wenting Liao, Jingjing Xu, Feifei Wang, Qing Liao, Shengnan Li, Minmin Liao, Yu Li, Linshang Rao, Jinquan Lin, Jianyuan Shi, Zejun You, Wenlong Zhong, Xinrong Liang, Hao Han, Yan Zhang, Na Tang, Aixia Hu, Hongyi Gao, Zhiqiang Cheng, Li Liang^&, Weimiao Yu^&, Yanqing Ding^&. Hybrid AI-assistive diagnostic model permits rapid TBS classification of cervical liquid-based thin-layer cell smears. (2021). Nature Communication. (IF = 16.6)


• Laurent GOLE^#, Joe YEONG^#, Jeffery Chun Tatt LIM, Kok Haur ONG, Aye Aye THIKE, Yong Cheng POH, Sidney YEE, Jabed Iqbal, Bernett LEE*, Wanjin HONG^&, Weimiao YU^& and Puay Hoon TAN^&. Quantitative stain-free imaging and digital profiling of collagen structure reveal diverse survival of triple negative breast cancer patients. (2020), Breast Cancer Research. (IF = 5.9)


• Shyi-Chyi Wang, Tricia Yu Feng Low, Yukako Nishimura, Laurent Christophe Michel Gole, Weimiao Yu^&, Fumio Motegi^&, Cortical forces and CDC-42 control clustering of PAR proteins for C. elegans embryonic polarization. (2017).  Nature Cell Biology. (IF = 28.2)


• Chiara Malinverno, Salvatore Corallino, Fabio Giavazzi, Martin Bergert, Qingsen Li, Marco Leoni, Andrea Disanza, Emanuela Frittoli, Amanda Oldani, Emanuele Martini, Tobias Lendenmann, Gianluca Deflorian, Galina V. Beznoussenko, Dimos Poulikakos, ONG Kok Haur, Marina Uroz, Xavier Trepat, Dario Parazzoli, Paolo Maiuri, Weimiao Yu, Aldo Ferrari, Roßberto Cerbino, Giorgio Scita. Endocytic re-awakening of motility in jammed epithelia. . (2017) Nature Materials. (IF = 41.2)


• Adam Cliffe^#, David P. Doupé^#, HsinHo SUNG, Isaac Kok Hwee LIM, Kok Haur ONG, Li CHENG,  Weimiao YU^&. Quantitative analysis of complex individual cell behaviors in highly coordinated in vivo collective cell migration. (2017). Nature Communication. (IF = 16.6)


• View full list of publications here