Every year, pathologists in the NHS manually examine millions of slides from bowel cancer screening biopsies. A new AI tool called COBIx aims to pre-screen those slides, automatically sorting them into three categories: normal, abnormal non-neoplastic, and abnormal neoplastic (cancerous or pre-cancerous). This matters because the NHS is short of pathologists, and the number of bowel biopsies is rising. Currently, every slide—normal or not—requires a consultant’s full review. That delays diagnosis for patients with serious disease. COBIx would remove normal slides from the pathologist’s workload entirely and flag abnormal cases as urgent or non-urgent, so the sickest patients are reviewed first. If this multi-site validation succeeds, the tool could speed up treatment decisions across the NHS. Patients with serious bowel disease would get faster diagnoses, while pathologists could focus their time on cases that need expert judgment. The algorithm has already been tested on 1,700 unseen slides, achieving an area under the receiver operator curve of 0.96–0.99—performance comparable to human pathologists. The next iteration, adding individual cell recognition, aims to improve accuracy for non-neoplastic disease. The study will also generate the safety and health economic data needed for regulatory approval.
View original technical description
Currently all endoscopic large bowel biopsies are examined manually, mostly by consultant pathologists. This proposal aims to use AI pre-screening of slides to reduce workload and improve workflow. The co lon and rectal bi opsy reporting AI tool (COBIx) we have produced screens haematoxylin and eosin stained biopsies segregating them into normal, abnormal neoplastic and abnormal non-neoplastic groups. We propose using this tool to remove normal slides from the pathologist workload and use the severity of disease detected to triage abnormal cases into urgent and non-urgent groups. As a result pathologist time is focussed on disease, meaning patients with serious disease are prioritised for immediate review. For all groups patient's samples will be reported more quickly, with greater priority given to serious disease, leading to faster treatment decisions. This helps deliver the NHS Long Term Plan's [1] aims to diagnose cancer earlier and to support pathologists, clinicians and patients by concentrating health resources on those patients that require treatment. COBIx has been developed as part of the PathLAKE project. The algorithm has been trained on ~3,400 whole-slide images (WSI) and tested on 1,700 unseen slides delivering an area under receiver operator curve of 0.96-0.99. These results demonstrate highly effective segregation of slides into the three categories described, with high sensitivity and negative predictive values comparable to human pathologists. The next iteration which includes individual cell recognition is now being introduced to deliver greater accuracy for the identification of non neoplastic disease. We have made careful provision for IVDCE and UK-CA clearance, sharing with BSI and the MHRA our approach to the development of this tool. The multi-site study outlined in this study plan will provide the efficacy and safety data needed for regulatory approval, as well as key health economic data indicating the impact the technology will have in routine practice.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know