Abstract
The interpretation of dental radiographs often varies due to the subjective nature of the
diagnostic process, leading to inconsistencies in diagnoses and treatments. To address
this challenge, this research focuses on the development of an advanced model for dental
disease detection. This model, utilising crowdsourced annotations from radiographic
images, aims to assist dental practitioners in identifying basic dental abnormalities from
radiographs.
Introducing Crowd-YOLO, an innovative adaptation of the object detection model
YOLOv3, this approach incorporates Bayesian classifier combination to effectively measure
data uncertainty and enhance accuracy. Crowd-YOLO distinguishes itself by extracting
reliable information from the majority of non-expert annotations and very
limited annotation from the expert, employing a confusion matrix to evaluate each annotator’s
capability in diagnosing various diseases. This methodology presents a novel
approach to mitigating discrepancies among different annotators.