Interobserver variability in the interpretation of Papanicolaou tests using the Hologic Genius Digital Diagnostics System: detailed analysis of cytology verification in a large academic laboratory
Haley Corbin, Jonee Matsko, Taylor Charity, Amy Colaizzi, Haley Brown, Jensen Bendig, Lakshmi Harinath, Beth Clark, Esther Elishaev, Rohit Bhargava, Liron Pantanowitz, Chengquan ZhaoAbstract
Objectives
Whole-slide digital imaging technology paired with artificial intelligence is transforming the practice of cervical cancer screening. We aimed to evaluate the performance of the Hologic Genius Digital Diagnostics System (HGDDS) for detection of neoplasia and to investigate interobserver variability among users who have undergone HGDDS training.
Methods
A total of 110 ThinPrep Papanicolaou tests were selected, with original diagnoses that included 47 cases of low-grade squamous intraepithelial lesion or worse (LSIL+); 4 cases of unsatisfactory findings; and 59 cases negative for intraepithelial lesion or malignancy (NILM), including 23 cases with microorganisms. All cases were reviewed by 16 cytologists or cytopathologists using the HGDDS. Interobserver variability and correlation with original diagnoses were assessed.
Results
The HGDDS showed greater than 95% sensitivity, specificity, positive predictive value, negative predictive value, and accuracy for LSIL+ disease. The overall misinterpretation rate was 5.5%. In total, 3 cases had poor interpretation (≥7/16 misinterpretations), including a high-grade squamous intraepithelial lesion case with 2 adenocarcinoma responses, an LSIL response, and 4 unsatisfactory responses, as well as 2 NILM cases interpreted as unsatisfactory by most participants due to borderline cellularity. Of 23 cases with microorganisms, organisms were not present in the 60-image gallery for 4 cases. Sensitivity was highest for Trichomonas (93.8%) but was low for other microorganisms (58.9%-68.8%).
Conclusions
In a large academic cytology laboratory, HGDDS review by trained users demonstrated high concordance with adjudicated diagnoses for detection of LSIL+, supporting its effectiveness for identifying epithelial abnormalities. However, user experience with digital morphology is important. During early implementation, laboratories should monitor discordance and consider temporary dual review, with particular attention to borderline cellularity cases and microorganism detection.