Quantitative morphometry of the cervix using deep learning and geometric area estimation techniques.
Authors
Affiliations (4)
Affiliations (4)
- Department of Obstetrics and Gynecology, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India.
- Department of Electronics and Communication Engineering, Indian Institute of Information Technology, Design and Manufacturing, Chennai, India.
- Department of Biostatistics, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India.
- Department of Radiation Oncology, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India.
Abstract
Estimation of the actual cervix area observed during the colposcopy is a critical step towards accurate diagnosis, as hidden areas may lead to malignancy in the future. In the present study, an automated algorithm was developed that uses a deep learning segmentation model combined with geometric area estimation techniques to evaluate the percentage of the cervix visualized during colposcopy by the clinician in real-time. The colposcopy images from 110 patients were examined with standardized Cusco bivalve specula and were manually annotated to train the segmentation model for YOLOv8s-seg training (100 epochs, batch size 16, learning rate 0.001, COCO pretrained weights). After segmentation, four computational methods: Grid-based method, pixel-based method, convex hull-based method, and a hybrid weighted model integrating these three metrics via weighted averaging, were applied to estimate the percentage of cervix visualized. The estimated cervical visualization scores generated by the proposed methods were validated against the expert visualization scores independently by three gynecologists, who provided independent visual assessments as the clinical reference standard. Agreement was evaluated using Pearson correlation (r), coefficient of determination (R<sup>2</sup>), and P values. The hybrid model demonstrated a higher correlation with the expert consensus with the Pearson's coefficient r = 0.7231, R<sup>2</sup> = 0.52288, P < 0.0001, residual sum of squares = 36.34. Individual methods gave a moderate performance with grid-based (r = 0.497, R<sup>2</sup> = 0.247), pixel-based (r = 0.503, R<sup>2</sup> = 0.253), and convex hull-based (r = 0.108, R<sup>2</sup> = 0.012, P = 0.261). Hybrid fusion achieved synergistic improvement over single methods with medians closely matching expert assessment (~85%-90%). This automated hybrid approach demonstrated significant correlation with the clinician grading by addressing a combination of key different image quality metrics during colposcopy and proposing an efficient and an empirical tool for real-time estimation of cervix visualization.