Projects using AI in detection and diagnosis
Prostate biopsy tissue reviewed using Paige Prostate AI (tissue suspicious of cancer is dark pink)
The ARTICULATE Pro Study
Men being investigated for prostate cancer usually have a prostate biopsy. Small samples of prostate tissue are sent to a Histopathology / Cellular Pathology laboratory for processing where the biopsy tissue is placed on glass slides. The slides generated are scanned, and then viewed by a pathologist on a digital screen in order to assess if cancer is present or not and how aggressive it looks (Gleason Grade). This helps clinicians to make treatment decisions for patients.
By using digital image-based diagnostics rather than glass slides, pathologists can create a more efficient and effective service. Digital pathology also enables the use of AI (artificial intelligence) diagnostic systems. For instance, much of the screening, grading and measuring could be done by pathologists with the assistance of AI tools (computer assisted technology) particularly in prostate cancer diagnostics, where there are already regulatory approved tools available.
One such tool is Paige Prostate – this is a clinical-grade, computer assisted diagnostic software system that helps pathologists with detecting, grading and measuring prostate tumours in biopsies obtained from patients at risk of prostate cancer.
The clinical software assists pathologists with:
- Highlighting areas suspicious of cancer
- Gleason Grading (how aggressive the tumour looks and therefore might behave)
- Measurements of the amount of tumour present.
The system is tested and proven to help pathologists diagnose more accurately and more efficiently and is approved by regulatory bodies to be used in clinical practice.
We are using the Paige Prostate diagnostic system in a study, called ARTICULATE PRO led by Professor Clare Verrill, to assist prostate biopsy reporting across three NHS sites in Oxford, Bristol and Coventry and Warwick and to evaluate the technology in use by pathologists. This study is an “Evaluation of Clinical Care” (service evaluation) and is funded by the NHS Accelerated Access Collaborative (NHS AAC) and NHSx.
Robustness and reproducibility for AI learning in biomedical sciences: RENOIR
AI techniques are revolutionising our capability to analyse complex and large scientific datasets. However, lack of reproducibility and poor generalisation are two significant challenges affecting AI publications. In response to these critical issues, Dr Alessandro Barberis and colleagues developed RENOIR, a software implementation of a generalized analytical framework using resampling for training and testing, ensuring robust and reproducible machine learning model development.
RENOIR adopts standardised pipelines and automated generation of transparent reports, enhancing reproducibility. Its object-oriented implementation supports straightforward expansion of currently supported settings. Moreover, clear interactive plots and summary tables integrated in the reports facilitate exploratory analysis and make our software accessible to a broader audience, beyond computer science and bioinformatics experts.
This work is a collaborative effort with Professors Francesca Buffa and Hugo Aerts, and the software is an initial implementation of the framework they used in a few studies, including:
- "A machine learning and directed network optimization approach to uncover TP53 regulatory patterns" by Triantafyllidis et al., published in iScience;
- "A role for SETD2 loss in tumorigenesis through DNA methylation dysregulation" by Javaid et al., published in BMC Cancer.
The team are excited to share that an initial implementation of this framework is available as an R package on GitHub.