ABSTRACT
OBJECTIVE: The present study aims to quantify clinicians' perceptions of oral potentially malignant disorders (OPMDs) when evaluating, classifying, and manually annotating clinical images, as well as to understand the source of inter-observer variability when assessing these lesions. The hypothesis was that different interpretations could affect the quality of the annotations used to train a Supervised Learning model. STUDY DESIGN: Forty-six clinical images from 37 patients were reviewed, classified, and manually annotated at the pixel level by 3 labelers. We compared the inter-examiner assessment based on clinical criteria through the κ statistics (Fleiss's kappa). The segmentations were also compared using the mean pixel-wise intersection over union (IoU). RESULTS: The inter-observer agreement for homogeneous/non-homogeneous criteria was substantial (κ = 63, 95% CI: 0.47-0.80). For the subclassification of non-homogeneous lesions, the inter-observer agreement was moderate (κ = 43, 95% CI: 0.34-0.53) (P < .001). The mean IoU of 0.53 (±0.22) was considered low. CONCLUSION: The subjective clinical assessment (based on human visual observation, variable criteria that have suffered adjustments over the years, different educational backgrounds, and personal experience) may explain the source of inter-observer discordance for the classification and annotation of OPMD. Therefore, there is a strong probability of transferring the subjectivity of human analysis to artificial intelligence models. The use of large data sets and segmentation based on the union of all labelers' annotations holds the potential to overcome this limitation.
