Related Experiment Video For Artificial intelligence
Updated: Sep 13, 2025
Introduction of an Integrated Pathology Image Management, Artificial Intelligence, and Reporting System
Published on: July 11, 2025
Real-time detection of premalignant cervical lesion using Artificial Intelligence (AI) model in multispectral imaging
A Keerthana1, Arpitha Anantharaju2, P T Mohammed Ansar1
1Department of Electronics and Communication Engineering, Indian Institute of Information Technology, Design and Manufacturing, Kancheepuram, Chennai, India.
Background:
Cervical cancer is the fourth most common cancer among women globally. The early diagnosis of cervical cancer can significantly improve prognosis and effective prevention. The accuracy of the cervical examination varies due to inter-clinician variability.
Objective:
This work is aimed to develop a real-time and static AI model integrated in Multispectral imaging System (GynoSight) with Graphical User Interface to assist the clinician in detecting abnormal blood vessels, acetowhite uptake, and iodine negative regions during the cervix screening.
Methods:
In this study, the dataset is acquired from the Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry between February 2024 to December 2024. The colposcopy images were acquired from the subjects between the age group of 22-65 years after smearing normal saline, 3 % acetic acid, and Lugol's Iodine. The preprocessing steps are applied on all images and the dataset is framed with 609 images, 196 normal saline-smeared cervix images,212 acetic acid-smeared cervix images, and 201 Lugol's iodine-smeared cervix images which were used to train and validate the model. The EfficientNet architecture AI model was implemented to identify the atypical blood vessel, dense acetic acid uptake and negative uptake of Lugol's iodine. The transvaginal imaging probe (GynoSight) was used to perform real-time detection of region of interest using a trained TensorFlow model.
Results:
The proposed model achieved an accuracy of 86.67% for the identification of Iodine negative regions and an F1 score of 90% and mAP (50%) of 91.1%. The updated proposed model implemented for the detection of the acetic acid uptake was reported with an improvement of accuracy from 57% to 85.71% after applying segmentation algorithm which indicates the reliability and robustness of the system. Furthermore, the detection of atypical blood vessels, acetowhite region, and Iodine negative uptake regions by the AI system is compared with clinician interpretation followed by the biopsy results.
Conclusion:
The AI-assisted real-time detection system may assist the clinician during screening, which will reduce the number of unnecessary biopsies for the patients, reducing the inter-clinician variability.
