Estimation and Comparison of Cortical Thickness Index and Canal-to-Calcar Ratio Using Manual Method and Deep Learning Method

Abstract

Manual calculation of the cortical thickness index (CI) and canal-to-calcar ratio (CC) using radiographs has been widely used. The purpose of this study was to investigate the difference between manual gold standard and automatic calculation based on deep convolutional neural networks (CNNs) of the proximal femur. We obtained institutional review board approval to utilize previous radiographs for the study and the radiograph images were used to train CNN architecture. The calculation experiment of a dataset of 136 images of the proximal femur to estimate CI and CC was performed using a trained CNN architecture (Automatic). Also, manual segmentation method (Manual) to calculate CI and CC was conducted using the standard protocol by two experts as a reference for the results comparison. The mean values of the Manual and Automatic calculation of CI for the proximal femur were 0.56 and 0.54, respectively, showing a statistically significant difference (p = 0.035). Significant difference (p < 0.001) was also seen in the calculation of CC by Manual and Automatic method resulting in 0.42 and 0.47, respectively. In addition, the automatic method showed far better results in terms of calculation speed (less than 30 s per single image). Therefore, we suggest that the manual method be carefully considered while planning a surgery.