Polyp Classification

Colorectal cancer is the second most common cancer and the second most common cause of cancer death in Norway. Organised bowel cancer screening is implemented in several countries and is shown to reduce both incidence as well as mortality rates significantly. Norway will begin screening in 2021 but implementation will require more endoscopists and pathologists, and in particular specialised gastrointestinal (GI) pathologists with experience from histological polyp examinations. Norway has few endoscopists, few pathologists and even fewer specialised GI-pathologists with the required experience. Automated risk classification of colon polyps will drastically reduce the workload for the pathologist, improve objectivity and allow for the increased throughput that is required to implement a bowel cancer screening programme in Norway. We aim to develop an automated histology classification system for bowel polyps using deep learning.

A faecal occult blood test (FOBT) is commonly used to identify patients that should undergo a colonoscopy to examine the bowel for polyps. When polyps are identified during a colonoscopy procedure, these are removed and examined by a specialised GI-pathologist. Further follow-up is decided based on the pathologist’s examination, who classifies the polyps according to histological type, where the different types are associated with a low or a high risk of developing into invasive cancer. Patients with high-risk polyps are scheduled for more extensive and more frequent follow-up investigations than patients with low-risk polyps. The vast majority of the identified polyps are low-risk adenomas, which very seldom develop into adenocarcinomas over time and require no further treatment when removed.

Interobserver agreement in the reporting of polyp pathology is suboptimal. This situation is not ideal for the patient, who may end up with suboptimal treatment, and underpins the need for more objective guidelines and methods. Furthermore, pathologists are a scarce resource with a significant workload, and are often the bottleneck in most countries’ treatment pathways.

We aim to develop an automated histology classification system for bowel polyps using deep learning that classifies a polyp’s pathology according to whether it has a histology type associated with a definite low risk or a high risk for developing into cancer. With such a system properly implemented, the pathologist will have to examine only the identified potential high-risk polyps (at most 10%), whereas the remaining definite low-risk polyps (90% or more) can be left without further investigation. The approach will drastically reduce the workload for the pathologist, improve objectivity and allow for the increased throughput that is required to implement a bowel cancer screening programme in Norway.

The project will be implemented in close collaboration with researchers from the University of Oxford, University College London (UCL) and Cheltenham General Hospital, where the above-mentioned leading GI-pathologists from UCL and Cheltenham General Hospital will be responsible for the data and the data quality used for training the deep learning model. As the pathological examinations of polyps have been carried out for more than ten years as part of the screening programme in the UK, the data availability is unlimited for practical purposes. We have agreed to utilise a discovery dataset consisting of 6800 polyps from 2800 patients diagnosed at UCL, while a dataset consisting of 6000 polyps diagnosed at Cheltenham General Hospital will be used for an independent validation of the method. The widespread interest in this system ensures a broad application when it is successfully developed and implemented in 2023.

In part I, we aim to develop a deep learning-based histology classification system for bowel polyps using the pathologists’ histology classification as ground truth. This system will identify polyps with an associated low-risk of developing into cancer and thereby reduce the pathologists’ workload significantly as this group constitutes the majority of polyp diagnoses. Furthermore, the system will support the pathologists in their polyp examination by providing classification suggestions that may improve the consistency and accuracy of polyp diagnostics. First, we will develop a system that automatically identifies tubular adenomas without high-grade dysplasia. This histological subgroup constitutes more than 50% of all polyp diagnoses and is associated with a low risk of developing into cancer. Robust automatic identification of polyps belonging in this group will eliminate a major portion of the pathologists’ polyp examination workload. By using our competence and experience on deep learning in the analysis of scanned tissue sections, we consider this task highly feasible. We will then add normal and hyperplastic polyps to the definition above. These histological subtypes constitute more than 75% of all polyp diagnoses and are also associated with a low risk for developing into cancer. An absolute requirement in the identification of low-risk polyps is that none of the high-risk polyps are erroneously classified as low-risk. The next phase is more complex and includes the development of a deep learning-based system for the histology classification of the remaining polyps (after the simple cases have been identified). The overall aim is to automatically identify the definite low-risk polyps (90% or more) that are associated with a low risk for cancer development and to leave the remaining polyps for the pathologists to examine thoroughly. The approach will drastically reduce the workload for the pathologist, improve objectivity and allow for the increased throughput required to implement a bowel cancer screening programme in Norway at full scale. Even if not all histology classes among the low-risk polyps can be automatically detected, automatic classification of e.g., 50-70% of all polyps without errors is still a great contribution to polyp diagnostics. In part II, we will introduce the deep learning system to the real gold standard, patient outcome, and combine histological classification and patient outcome in the development of a deep learning system for the estimation of risk of colorectal cancer development given the observed bowel polyps. We will identify and include patients enrolled in the UK bowel cancer screening programme that later presented with colorectal cancer and use this information together with the histological type for labelling samples during training. Polyps from patients who later developed colorectal cancer will represent increased cancer risk compared to polyps from patients who did not, all else being equal, with this approach. Furthermore, we will manually re-evaluate the histological diagnosis of the polyps in this patient group to verify the diagnosis and acquire new insight about the characteristics of these polyps. Some histological classes are particularly difficult to distinguish, even for expert GI-pathologists. We postulate that the approach in part II of the project may reveal knowledge that can be used to improve the consistency and accuracy of these difficult diagnoses.