It’s done!
Our study is published, here is an introduction.
Each year, nearly 30,000 pre cancers and 3,000 cases of cervical cancer are reported in France, and 1,100 women still die from it. Although human papillomavirus (HPV) screening, the virus responsible for cervical cancer, effectively detects the presence of this infectious microorganism, it does not allow, however, to distinguish with certainty between harmless transient infections and precancerous or cancerous lesions.
This low specificity results in an increase in consultations and an increase in the number of colposcopies (estimated at 70%). This examination, which involves observing the cervix using a microscope after applying a dye, reveals precancerous or cancerous lesions.
However, more than 40% of the colposcopies performed do not detect any abnormalities. This situation leads to overdiagnosis and overtreatment, particularly by conization. This is a procedure that removes part of the cervix and can have repercussions on fertility and pregnancy.
An AI tool to reduce unnecessary conizations and biopsies
The results of the artificial intelligence tool Cervital®, designed by Dr. Joseph Monsonego, president of the col-HPV commission of the French National College of Gynecologists and Obstetricians (CNGOF), have just been published in the journal Computers in Biology and Medicine. This work confirms the promising and reliable nature of this tool, which optimizes the prediction, evaluation, and instant recognition of normal cervixes as well as precancerous lesions induced by HPV.
The tool in question is a diagnostic risk prediction model based on deep learning. It was trained, tested, and evaluated using a database of 30,000 patients. For each of them, medical history information was available (cytological and/or HPV screening results, histological analyses of biopsies or conizations, as well as imaging data collected during consultations over more than 20 years). The model’s performance evaluation was then based on a sub-population of 6,356 patients who, after an abnormal screening result, underwent conization or cone biopsy (a procedure that involves removing a cone-shaped fragment of the cervix to analyze the cervical cells in depth and identify any precancerous or cancerous lesions).
The final diagnosis of these patients, considered an indisputable reference, served as the “gold standard” for identifying CIN2+ (high-grade lesions with a risk of evolving into cancer). The overall performance, discrimination, and calibration of the model were compared to the opinion of an expert clinician on colposcopy.
Artificial intelligence outperforms an experienced practitioner’s expertise
In conclusion, at this stage, the artificial intelligence model outperforms the expertise of an experienced practitioner, with a performance superior by 10 points for detecting precancerous lesions (CIN2+). “Thus, many conizations could be avoided without missing any real cases of CIN2+”, explain the authors, which confirms the potential clinical utility of deep learning models to reduce unnecessary conizations or biopsies.
One can even imagine that this difference could be even more significant when the tool is used by a less experienced practitioner, thus reinforcing its interest in improving diagnostic accuracy and standardizing care throughout the territory.
The next step
This digital solution will now undergo external (by testing it on data from other sources and different populations), multicenter, and prospective validation. By making this tool accessible to everyone, researchers hope, artificial intelligence promises to accurately assess the risk of precancerous lesions and cervical cancer, improve pre-cancer detection, and assist the practitioner in diagnosis and decision-making. Moreover, besides the time-saving offered to doctors, this device will help alleviate patients’ concerns.
Source: https://www.sciencedirect.com/science/art…
To receive the study by mail, please fill out the form below: