Short summary

Colposcopy (the examination of the cervical to diagnose cervical dysplasia) is a subjective examination and very dependant on the colposcopists performing the exam. With new developments within AI technology we wish to examine whether we can build an effective AI algorithm to provide a more objective approache and evaluate potential pre-malignant/malignant changes of the cervix and improve the diagnostics of cervical dysplasia and cervical cancer.

Rationale

In Denmark, screening for cervical cancer is free of charge and offered to women between the ages of 23 and 64. Women with an abnormal screening test, indicating potential cervical pre-cancerous lesions, are referred for further diagnostics through colposcopy with cervical biopsies, which is performed by a gynecological specialist.

Colposcopy with biopsy is the diagnostic standard worldwide for cervical dysplasia. Different patterns on the cervix are identified during colposcopy which could indicate cervical intraepithelial neoplasia (CIN) (aceto-white, lesion margins, mosaic vessel patterns, punctuations, and atypical vessels). Based on the visual findings colposcopist will evaluate potential areas on the cervix for biopsy. Biopsies are sent for histological evaluation; this histological diagnosis dictates further follow-up and/or treatment. Women diagnosed with high-grade cervical lesions (CIN2 or worse, referred to as CIN2+) typically undergo conization. However, women with future fertility wishes are managed conservatively when diagnosed with CIN2, due to increased risk of preterm delivery in pregnancy after conization. Colposcopy is a subjective procedure and is highly dependent on the experience of the colposcopist performing the examination, with high inter-observer and intra-observer variability. Different studies have revealed variable sensitivities of colposcopy as low as 55-57%. Previous studies have shown that taking additional random biopsies from areas with no visible lesions can significantly improve the sensitivity of colposcopy. Therefore, the Danish national guidelines on colposcopy recommend that four biopsies are always taken, regardless of visible lesions. Receiving an abnormal screening result and undergoing colposcopy can cause women to experience adverse psychological outcomes, and additional biopsies can lead to increase pain, discharge, and bleeding. Additional biopsies in all women also provides an increased workload for pathology departments who analyze all specimens. The national cervical cancer screening program is currently changing from cytology-based screening to HPV-based screening. HPV-screening has been found to be more sensitive and it is expected to detect smaller lesions than cytology testing alone.(26) As a result of these changes, more women will be referred for colposcopy with potentially less visible lesions to the naked eye. Different technologies have been developed to aid colposcopists by providing a more objective approach to colposcopy. The digital colposcope, DYSIS, is the most widely used of these new technologies. From 2017 to 2021 we performed a large multicenter study evaluating this technology. However, we found that the guidance provided by the DYSISmap was not sufficient and only provided a sensitivity of 42% for detecting high-grade lesions (CIN2+). During this project, we were able to collect a unique photographic and video material of cervical dysplasia and biopsy placement. Utilizing artificial intelligence (AI) - especially deep learning techniques-has received great attention in recent years and AI could have a pivotal role in the future of non-invasive diagnosis of diseases such as cervical cancer. This has the potential to increase the diagnostic accuracy of the colposcopy examination by guiding the colposcopists to improve biopsy placement and reduce the need to four biopsies in all women. This would provide women with better and timely diagnostics, leading to better planned follow-up, decrease discomfort during examination, potentially reducing number of visits and number lost to follow-up. This will also lower the number of unnecessary treatments and the experience of illness and anxiety related to a positive screening result. This in turn will lower the workload for gynecologists and pathologist, due to fewer follow-ups and fewer biopsies. It would also prevent cervical cancer cases, reducing the societal cost of oncology treatment.

Objective: The aim of the current research project is to generate and test clinically an artificial intelligence algorithm to aid colposcopists in choosing the correct area of the cervix to biopsy in women referred for colposcopy due to a positive screening test for cervical dysplasia.

Description of the cohort

Previously collected images from colposcopy examinations performed at Randers Regional Hospital between 2017-2021.

Data and biological material

Colposcopy images.

Collaborating researchers and departments

Unit at Applied AI and Data Sciences, Mærsk Mc-Kinney Møller Institute, University of Southern Denmark

• Professor Esmaeil Nadimi

Department of Gynecology and Obstetrics, Randers Regional Hospital

• Associate Professor Pinar Bor

Danish Centre for Clinical Artificial Intelligence (CAI-X), University of Southern Denmark

Department of Gynecology and Obstetrics, Odense University Hospital

• Professor Lone Kjeld Petersen