Skip to content

Research at St Andrews

Augmenting TNM staging with machine learning-based immune profiling for improved prognosis prediction in muscle-invasive bladder cancer patients

Research output: Contribution to conferencePoster

Open Access permissions



Nicolas Brieu, Christos Gavriel, David James Harrison, Guenter Schmidt, Peter David Caie

School/Research organisations


Background: Muscle-invasive bladder cancer (MIBC) is a highly aggressive disease whose clinical reporting is based on TNM staging. A more accurate and personalized prognosis could be achieved by profiling the immune contexture alongside clinical TNM staging. This study reports on features captured from the labelling of cells for CD3, CD8 and PD-L1 expression, within both the tumor and stroma of MIBC patients. This data is distilled to identify a novel immune-based prognostic signature which augments TNM staging. Methods: An image analysis solution was developed to quantify the density of cell populations across immunofluorescence (IF) labelled whole slide images from 105 MIBC patients with known survival data. A regression random forest was used for the detection of nuclei using the Hoechst channel [Brieu et al., ISBI2017] and a convolutional neural network employed for the segmentation of the tumor from the stroma using the pan-cytokeratin channel [Brieu et al., SPIE2018]. The CD3, CD8 and PD-L1 channels were used to classify the cells and calculate their proportion within either the tumor or stroma regions. A decision tree of depth two was trained on these proportions as well as cross-validated. More explicitly, the patients were recursively partitioned during training to maximize at each node their survival difference, leaves showing low survival difference (p-value>0.5) being finally merged. Results: The method yielded a decision tree which stratified patients into three groups utilizing only two parameters: the proportion of CD8(+) cells in the stroma and the proportion of PD-L1(+) cells across the whole tissue, which were negatively and positively correlated with cancer-specific death respectively. This method was used to replace TNM stages 1 to 3 while retaining the original stage 4 stratification. Testing for survival curve differences showed that this combined system yielded a higher prognostic value (Chisq=41.5, p-value=5.0x10-9) than the standalone TNM staging system (Chisq=34.9, p-value=1.25x10-7). Conclusions: Our results suggest that immune profiling derived from image analysis provides additional prognostic value to TNM scoring for MIBC.


Original languageEnglish
Publication statusPublished - 8 Oct 2018
Event43rd ESMO Congress 2018 - Munich, Germany
Duration: 19 Oct 2018 → …


Conference43rd ESMO Congress 2018
Period19/10/18 → …

Discover related content
Find related publications, people, projects and more using interactive charts.

View graph of relations

Related by author

  1. Automated tumour budding quantification by machine learning augments TNM staging in muscle-invasive bladder cancer prognosis

    Brieu, N., Gavriel, C., Nearchou, I. P., Harrison, D. J., Schmidt, G. & Caie, P. D., 26 Mar 2019, In : Scientific Reports. 9, 5174.

    Research output: Contribution to journalArticle

  2. Colorectal cancer outcome prediction from H&E whole slide images using machine learning and automatically inferred phenotype profiles

    Yue, X., Dimitriou, N., Caie, P., Harrison, D. & Arandjelovic, O., 18 Mar 2019, Proceedings of 11th International Conference on Bioinformatics and Computational Biology, BICOB 2019: Honolulu; United States; 18 March 2019 through 20 March 2019. Eulenstein, O., Al-Mubaid, H. & Ding, Q. (eds.). EasyChair, p. 139-127 11 p. (EPiC Series in Computing; vol. 60).

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

ID: 258788560