For the purposes of education and telemedicine, colposcopic appearance is commonly categorized using still images, which conceal the dynamic nature of acetowhitening. In this study, researchers evaluated the reliability of a colposcopic impression based on a single image taken 1 minute after acetic acid administration to that based on a series of 17 photos taken over 2 minutes. Roughly 5,000 colposcopies using the DYSIS colposcopic system were randomly split into 10 groups and given to 10 different, highly trained colposcopists. To determine whether a lesion was “normal,” “indeterminate,” “high grade,” or “cancer,” colposcopists first evaluated a single 2-D image taken at 1 minute and then evaluated a time-series of 17 subsequent photos. The ratings were matched to the histopathological diagnoses. The intra- and inter-rater reliability was estimated by having 5 colposcopists independently assess a random sample of 200 single images and 200 time series. Only 24.4% of 4,640 patients with sufficient images had their diagnoses confirmed by a single image visual assessment (11% of 64 cancers; 31% of 605 CIN3; 22.4% of 558 CIN2; 23.9% of 3,412< CIN2). Youden indices (sensitivity plus specificity minus one) for individual colposcopic examinations ranged from 0.07 to 0.24, indicating poor accuracy. Regardless of histology, using the time series led to a higher percentage of pictures being deemed normal. There was a lot of internal consistency between raters (weighted kappa = 0.64) and between raters (weighted kappa = 0.26). Even when a 17-image time series of the 2-minute acetowhitening process is shown, there is still substantial heterogeneity in the visual assessment of colposcopic pictures. To see if deep learning picture assessment can help with classification, the researchers are testing it out right now.
Source: sciencedirect.com/science/article/pii/S009082582200539X
