A novel SM-Net model to assess the morphological types of Sella Turcica using Lateral Cephalogram

Abstract

Abstract Objectives Deep learning (DL) models such as two pre-trained VGG models were explored and a novel SM-Net model is proposed to design an automated method for identifying different morphological types of Sella Turcica (ST). Further, all the models were compared based upon prediction results and evaluation metrics. Materials and Methods The lateral cephalogram dataset of 653 normal and patients with dentofacial were included and randomly divided into multiple subsets of training and testing data ratios. The manually labelled images encompasses pixel-by-pixel annotation of the Sella Turcica (ST) by dental specialists using an online labelling platform. The different image pre-processing techniques were employed to prepare the image dataset for convolutional neural network (CNN) modelling. The two pre-trained models Standard VGG-19 (SVGG-19), Optimised VGG-19 (OVGG-19) and a proposed SM-Net model were trained. These trained models extract Sella features by identifying an important region in the image and then classify Sella types based on pre-defined classes. Based on obtained training and validation accuracy graphs, we calculated pixel-wise IoU, mean IoU, and Dice coefficient to evaluate the performance of the models. Results The proposed SM-Net model shows significant training and prediction results compared to Standard VGG-19 (SVGG-19) and Optimized VGG-19 (OVGG-19). The mean IoU scores for Standard VGG-19 (SVGG-19), Optimized VGG-19 (OVGG-19) and SM-Net are 33.3%, 33.7%, 36.2% respectively and dice coefficients are 35.6%, 37.1, and 40.7% respectively. Conclusion The proposed fully-connected automated SM-Net model shows significant results towards detection and identification of morphological types of Sella Turcica (ST). Further work will be aimed to improve the accuracy of the selected model. Clinical Significance The proposed study will help dental experts and practitioners to pre-diagnose dentofacial anomalies associated with morphological features of Sella Turcica (ST).