IADR Abstract Archives
DMFT Index Evaluation Using Deep Learning on Panoramic Radiographs
Objectives: To evaluate the accuracy of a deep machine learning algorithm for assessing a DMFT index in panoramic radiographs of permanent dentition.
Methods: The dataset consisting of 2500 depersonalized panoramic radiographs were assessed and Decayed-(DT), Missing-(MT), Filled- (FT) and Unerupted-teeth (UT) [PR1] were marked independently by two trained and calibrated examiners. The intra-observer reliability Kappa values were 0.972 and 0.944 and the inter-observer reliability - 0.894. Inconsistencies were corrected by a third examiner. Python 3.8 programming language was used to develop machine learning algorithm to recognize DT, MT,FT and UT. EfficientNet B7 algorithm was applied to learn and evaluate DMFT values in panoramic radiographs. The performance of the model was evaluated using F1-score metrics and tested on unseen 257 images. The accuracy of DMFT score was assessed with the mean absolute error (MAE), additionally the area under the receiver operating characteristic curve (AUC) was determined and the F1 score was included. For each data item examined, one of the following values was assigned: true positive/negative (TP and TN) and false positive/negative (FP and FN).
Results: The number of TP values was 7658 (18.62%); TN - 32420 (78.84%); FP - 476 (1.16%); and FN - 566 (1.37%). The proposed method achieved a sensitivity of 0.9300, specificity of 0.9900, AUC of 0.9886, F-measure of 0.9360 and DMFT index MAE of 0.5136.
Conclusions: The current machine learning algorithm showed promising results for an automatic DMFT index assessment in panoramic radiographs. More data is needed for the further development of machine learning algorithm.
Methods: The dataset consisting of 2500 depersonalized panoramic radiographs were assessed and Decayed-(DT), Missing-(MT), Filled- (FT) and Unerupted-teeth (UT) [PR1] were marked independently by two trained and calibrated examiners. The intra-observer reliability Kappa values were 0.972 and 0.944 and the inter-observer reliability - 0.894. Inconsistencies were corrected by a third examiner. Python 3.8 programming language was used to develop machine learning algorithm to recognize DT, MT,FT and UT. EfficientNet B7 algorithm was applied to learn and evaluate DMFT values in panoramic radiographs. The performance of the model was evaluated using F1-score metrics and tested on unseen 257 images. The accuracy of DMFT score was assessed with the mean absolute error (MAE), additionally the area under the receiver operating characteristic curve (AUC) was determined and the F1 score was included. For each data item examined, one of the following values was assigned: true positive/negative (TP and TN) and false positive/negative (FP and FN).
Results: The number of TP values was 7658 (18.62%); TN - 32420 (78.84%); FP - 476 (1.16%); and FN - 566 (1.37%). The proposed method achieved a sensitivity of 0.9300, specificity of 0.9900, AUC of 0.9886, F-measure of 0.9360 and DMFT index MAE of 0.5136.
Conclusions: The current machine learning algorithm showed promising results for an automatic DMFT index assessment in panoramic radiographs. More data is needed for the further development of machine learning algorithm.