Classification of External Root Resorption Using Deep Learning-based Algorithm

Objectives: Early identification of external root resorption (ERR) is challenging due to the absence of clinical symptoms in most cases and is often detected incidentally during radiographic examination. Although Cone beam CT (CBCT) allows superior visualization of ERR, the clinician's skills resulted in subjective variability of the interpretation performance of CBCT images. A machine learning system combined with a pre-trained convolutional neural network (CNN) and a feature selection technique (FST) can potentially enhance the ability to identify ERR on CBCT images. This study aims to evaluate the performance of deep CNN models for the classification and segmentation of ERR on CBCT images.
Methods: Cone beam CT scans of 88 extracted premolars with simulated ERR were evaluated and labeled by an oral and maxillofacial radiologist and were defined as the reference dataset. Training and validation (80%) and test dataset (20%) were then established from the dataset. Random Forest algorithm combined with VGG16 deep CNN was employed to identify ERR according to different depths (0.5mm,1.0mm, 2.0mm). The Unet algorithm was applied to segment ERR lesions in all two-dimensional slices of the CBCT scans. The performance of the test dataset in the trained Random Forest+VGG16 and Unet models was evaluated in terms of accuracy, F1-score, and Intersection over Union.
Results: The classification accuracy using Random Forest+VGG16 improved from 78.6% to 81.1% following FST optimization. The segmentation accuracy measured by Intersection over Union was 63%. The volumetric measurement of the ERR was comparatively similar between the deep CNN system and manual segmentation.
Conclusions: Random Forest with VGG16 algorithm combined with FST has the potential to aid in the identification of ERR on CBCT images based on different depths. The Unet algorithm offers the potential for segmentation and volumetric ERR measurement.