Computational Digital Pathology Lab

Our vision is to prototype tailor-made computational medical image analysis solutions for digital pathology. Our scope includes image analysis and general bio-image quantification methods of pathology images such as from Hematoxylin-Eosin (H&E) and Immunohistochemistry (IHC) staining.

Our team focuses on the following topics in digital pathology and biomedical image analysis:

• Developing novel computational image processing algorithms for basic scientific research, drug discovery and other biomedical applications
  
• Developing practical industry-ready software packages for automated image or video analysis to enhance clinical workflow and efficiency
  
• Building mathematic models for classification and prediction using professionally curated Histopathological image databases
  
• Developing machine learning, AI-based, and deep learning solutions for big biological/biomedical imaging data
  

With recent advances in digital image technology and artificial intelligence, histology laboratories and standards are undergoing fundamental change. Automated quality assessment tools are essential for the integration of modern digital solutions into day-to-day clinical practice and the evaluation of image quality remains an ongoing area of research. In order to calculate an image's non-reference image quality score, functional estimators of perceptual image quality are necessary. Traditional estimators of image quality typically prove to have higher efficiency on natural images compared with histological images. While several tools have been previously developed for histological images, estimators of quality assessment in H&E images remain limited to focal accuracy. To address this issue, we have developed a dedicated solution specifically to enhance nuclear image quality assessment. At present, our software package has been validated by industry-renown digital image scanner manufacturers.

We have worked closely with a number of pathologists to develop a deep learning based model for automated Gleason Grading in prostate histopathological images which learns from professionally curated images. Based on expert annotations, the AI model is trained to differentiate Gleason Pattern 3, Gleason Pattern 4, Gleason Pattern 5, normal glands, stroma and other structures. At present, the AI has proven able to detect malignant regions at different magnifications and alert pathologists to certain, small malignant regions.  This automatic detection increases the accuracy, efficiency and consistency of the Gleason Grading and is designed to alleviate the manual burden of quantitative tumour assessment. It is also applicable to images scanned by scanners from different brands such as Akoya Biosciences, Olympus, Zeiss, KFBio and Leica. 

Moreover, the preliminary model we’ve built is capable of incremental learning, which means the model’s performance will improve gradually with the addition of new training data collected through its usage. With every interaction between the AI model and pathologists, new annotated data can be collected more efficiently to enhance the model and to benefit scientific research.

Mitotic count is a crucial clinical marker in breast cancer diagnosis. We’ve developed a pipeline for semi-auto mitotic figure detection in breast cancer histopathological images which were validated on three different data sets: SGH dataset, ICPR2012 and ICPR2014. 

Since Hematoxylin stains only the nuclei, where mitosis occurs, Haematoxylin and eosin channels are first separated by colour deconvolution and then analysed based on the intensity features of mitotic figures in the Hematoxylin channel. The three most discriminative features were selected to train a Naïve Bayes model to differentiate mitotic figures and background. By applying the trained model, regions that don’t possess mitotic figures can be filtered and only those that do are retained (~15% of the whole area) for the user’s evaluation. The output of the model assists the pathologists in locating ‘mitosis hot spots’ when counting the number of mitotic figures. This feature is applicable to images, including TMA images, of various size.

We have developed an automated image processing pipeline to detect and quantify multiple immunophenotypes in histological prostate tissue samples. Our pipeline integrates powerful image processing and machine learning methods to analyse histopathological tissue whole slice images (WSI). 

To develop our pipeline, we have stratified the patient cohort (N = 156) based on the patient Gleason score (GS, primary and secondary Gleason grade) annotated by the pathologists. We cut each patient’s tissue sample into 4 consecutive slices. The first piece is stained with H&E so that the pathologist can evaluate. The subsequent 3 sections were stained with UltiMapper I/O PD-L1, UltiMapper I/O PD-1 and UltiMapper I/O T-act kits. We use pipelines to quantify the immunophenotype along the four consecutive slices in terms of cell density (#cells/mm2). Then, we analysed the statistical significance of these immunophenotypes between patient groups and different tumour regions. For a variety of different patient groups, we used multiple immunophenotypes to obtain significant results, p <0.05. Our results combine basic overall cell density and neighbouring cell density, paving the way for more complex quantitative cell-cell interactions that can be used in the clinical environment for cell distribution pipelines.

Cervical cancer can be effectively treated if diagnosed early. Artificial intelligence has the potential to reduce pathologists’ workload, improve diagnostic speed and maintain high accuracy for cervical cytology screening. To improve diagnostic accuracy and speed, we developed an AI assistive diagnosis system to classify cervical liquid-based thin-layer cell smears in accordance with TBS criteria.

In this study, we trained our models with >81,000 samples retrospectively. We used YOLOv3 to develop the target detection model (mAP, Mean Average Precision, up to 0.8233), Xception algorithm and Patch-based classification to boost the detection target classification model (accuracy rates of 0.968 and 0.9192, respectively), U-net to segment the nucleus and accurately measure the gray value of the nucleus ( Mean intersection, mIOU reaches 0.8356. We integrated XGBoost and logical trees together with the above deep learning models based on optimized features given by the deep learning process, formulated a complete cervical liquid-based cytology smear TBS classification and diagnosis pipeline, and successfully established an artificial intelligence-assisted diagnosis system. The optimized system was validated with >34,000 prospectively samples collected from 11 medical Institutions.

• Zhu X, Li X, Ong K, Zhang W, Li W, Li L, Young D, Su Y, Shang B, Peng L, Yu W*, et al: Hybrid AI-assistive diagnostic model permits rapid TBS classification of cervical liquid-based thin-layer cell smears. Nature Communications 2021, 12(1):3541.


• Gole L, Yeong J, Lim JCT, Ong KH, Han H, Thike AA, Poh YC, Yee S, Iqbal J, Hong W, Lee B, Yu W, Tan PH. Quantitative stain-free imaging and digital profiling of collagen structure reveal diverse survival of triple negative breast cancer patients. Breast Cancer Research 22, 1-13, (2020)


• Tan EE, Hopkins RA, Lim CK, Jamuar SS, Ong C, Thoon KC, Koh MJ, Shin EM, Lian DW, Weerasooriya M, Lee CZ, Soetedjo AAP, Lim CS, Au VB, Chua E, Lee HY, Jones LA, James SS, Kaliaperumal N, Kwok J, Tan ES, Thomas B, Wu LX, Ho L, Fairhurst AM, Ginhoux F, Teo AK, Zhang YL, Ong KH, Yu W, Venkatesh B, Tergaonkar V, Reversade B, Chin KC, Tan AM, Liew WK, Connolly JE. Dominant-negative NFKBIA mutation promotes IL-1β production causing hepatic disease with severe immunodeficiency. Journal of Clinical Investigation, 10.1172/JCI98882, 2020


• F Liu, L Gole, KH Ong, L Li, D Young, J Zhao, H Rao, W Yu*, L Wei*. Quantitative Collagen Profiling of HCV Patient Signifies the Fibrosis Reversibility Post Sustained Virologic Response. Hepatology 70, 1006A-1007A. (2019)


• Copenhagen K, Malet-Engra G, Yu W, Scita G, Gov N, Gopinathan A. Frustration-induced phases in migrating cell clusters. Science advances 4 (9), eaar 8483-14. 2018


• Cliffe A, Doupé DP, Sung HH, Lim KH, Ong KH, Cheng L, Yu W*. Quantitative analysis of complex individual cell behaviors in highly coordinated in vivo collective cell migration. Nature Communication and in press. (2017)


• Malinverno C, Corallino S, Giavazzi F, Bergert M, Li Q, Leoni M, Disanza A, Frittoli E, Oldani A, Martini E, Lendenmann T, Deflorian G, Beznoussenko GV, Poulikakos D, Haur OK, Uroz M, Trepat X, Parazzoli D, Maiuri P, Yu W, Ferrari A, Cerbino R, Scita G. Endocytic re-awakening of motility in jammed epithelia. Nature Material, 1476-4660,http://dx.doi.org/10.1038/nmat4848, (2017)


• Diz-Muñoz A, Romanczuk P, Yu W, Bergert M, Ivanovitch K, Salbreux G, Heisenberg CP, Paluch EK. Steering cell migration by alternating blebs and actin-rich protrusions. BMC Biology, Sep 2;14:74. DOI: 10.1186/s12915-016-0294-x, (2016)


• Ong KH, De J, Cheng L, Ahmed S*, Yu W*. NeuronCyto II: An automatic and quantitative solution for crossover neural cells in high throughout screening. Cytometry Part A May 27. doi: 10.1002/cyto.a.22872, (2016)


• Gole L*, Ong KH, Boudier T, Yu W*, Ahmed S*. OpenSegSPIM: a user-friendly segmentation tool for SPIM data.  Bioinformatics. doi: 10.1093/bioinformatics/btw093, (2016).


• Ono Y, Yu W, Jackson HE, Parkin CA, Ingham PW. Adaxial cell migration in the zebrafish embryo is an active cell autonomous property that requires the Prdm1a transcription factor. Differentiation, http://dx.doi.org/10.1016/j.diff.2015.03.002, (2015)


• View full list of publications here