[Artificial Intelligence: accuracy for the diagnosis of precancerous lesions of the cervix].

INTRODUCTION

To compare the diagnostic sensitivity of artificial intelligence (AI) assisted videocolposcopy with standard videocolposcopy performed by specialist colposcopists.

METHODS

A descriptive retrospective cross-sectional study, 782 anonymized medical records from the Computerized System for Screening (SITAM) of women who underwent videocolposcopy with AI and colposcopy with common videocolposcopy performed by specialists, with their corresponding biopsies (gold standard) were analyzed. The relationship between the results of IA videocolposcopy and regular videocolposcopy and the results of biopsies was evaluated. The overall accuracy of each diagnostic procedure was calculated. The sensitivity and concordance of the results of AI videocolposcopy with the gold standard (biopsy) were determined.

RESULTS

A total of 395 patient records of patients with IA videocolposcopy and 387 with regular videocolposcopy were analyzed. The accuracy of results was 80% (IC 95%: 75-83%) in IA videocolposcopy and 65% (IC 95%: 60-69%) in regular videocolposcopy (p<0.001). Videocolposcopy results with IA and common colposcopy were significantly correlated with biopsy results, rs=0.75 vs. rs=0.57 respectively (p<0.001). The sensitivity of videocolposcopy with AI was 96% (95% CI: 94-98%), and 93% (95% CI: 89-95%) for regular colposcopy. The overall agreement of colposcopic impressions classified by videocolposcopy with AI and disease was higher than that of colposcopic interpretation by colposcopists (90% vs. 83%, Kappa 0.59 vs. 0.47, p<0.001).

CONCLUSION

The high diagnostic accuracy of AI videocolposcopy allows obtaining highly sensitive studies that help in the early detection of precursor lesions of cervical neoplasia.