Bone Diseases* ; Humans ; Radiography, Panoramic ; Radiologists
Abstract
Background: Development of a deep-learning-based diagnostic model requires labor-intensive work of medical image labeling. Moreover, when only a subset of the data is annotated with location labels, it limits the amount of data available for training models.
Methods: We (i) constructed a framework that enhances the diagnostic performance of jaw bone pathology by additionally utilizing network attention information, even with partially labeled data; (ii) introduced an additional loss in our training to minimize the discrepancy between network attention and attention labels; (iii) introduced a trapezoid augmentation method to maximize the utility of minimally labeled data. We collected panoramic radiographs from January 2019 to February 2021, for common jaw bone lesions and normal cases, totaling 716 data points. Two radiologists labeled the region-of-interest on each lesion’s periphery. To simulate varying availability of attention labels, we used subsets of 0%, 5%, 10%, 20%, 50%, and 100% attention-labeled lesion data for training, while class labels were consistently used across all data. ]
Results: Experiments on our dataset show that guiding the network attention with merely 5\% of attention-labeled data can enhance the accuracy of diagnostic models for jaw pathology from 92.41% to 96.57%. Moreover, the model performance was enhanced by the novel data augmentation methods introduced in the current study, which were more effective than previous combinations of data preprocessing and augmentation methods (accuracy, 99.17%).
Conclusion: The results support that the proposed framework contributes to fine-grained diagnosis using minimally labeled data.