Table 1.

Summary of recently published deep learning-based methods for mandibular canal segmentation in CBCT scans.

Author, Study and Year of Publication	Technique	Type of Dataset and FOV	No. of CBCT Scans		Contributions	Limitations
			Training + Validation	Test
Kwak et al. [19], 2020	Thresholding-based teeth segmentation + 3D UNets	Private, Full View	82	20	Employed 2D and 3D Deep Learning models and demonstrated the superior performance of 3D UNets	Limited performance in terms of Mean mIoU
Jaskari et al. [20], 2020	3D Fully Convolutional Neural Networks (FCNNs)	Private, Medium View	509	128	The study utilized a large number of CBCT scans to train 3D FCNNs and achieved an improved performance.	Overall achieved performance for left and right canal was limited in term of Dice score
Faradhilla et al. [21], 2021	Residual FCNNs + Dual auxilary Loss functions	Private, 2D view	NA	NA	The study exploited Residual Fully Convolutional Network with dual auxiliary loss functions to segment the mandibular canal in parasagittal 2D images and reported promising results in terms of dice score	Requires manual input from dentists to generate the 2D parasagittal views from CBCT. Study provides no information about the CBCT scans used for the experimentation
Verhelst et al. [25], 2021	3D UNet trained in two phases	Private, Medium View	160	30	Trained 3D UNet in two phases, i.e., before and after the deployment, to achieve promising performance.	Requires an extensive effort to train the model and inputs from experts are needed to improve its performance of the model.
Widiasri et al. [22], 2022	YOLOv4	Private, 2D view	NA	NA	The study utilized YOLOv4 for mandibular canal detection in 2D coronal images and achieved significantly higher detection performance.	The study used 2D coronal images which need manual input to generate from CBCT scans. The technique just provides the bounding box around the canal region, which lacks the exact boundary information which can be obtained from segmentation.
Lahoud et al. [26], 2022	Two 3D UNets, one for coarse segmentation and other for finetuning on patches	Private, Mixed FOVs	196	39	Adjusted to the variability in Mandibular Canal shape and width by using voxel-wise probability approach for segmentation	The scheme requires an extensive effort to train the models and evaluate performed on a limited private dataset does not prove the generalization
Cipriano et al. [27], 2022	Jaskari et al. [20], 2020	Public, Medium view	76 with Dense annotation	15 with Dense Annotations	The first publicly released annotated dataset and source code, validated their dataset on three different existing techniques	Utilized the existing segmentation methods, with no contribution in terms of technique novelty.
Cipriano et al. [28], 2022	3D CNN + Deep label propagation technique	Public, Medium view	76 with Dense annotation+256 with Sparse Annotations	15 with Dense Annotations	Combined 3D segmentation model trained on the 3D annotated data and label propagation model to improve the mandibular canal segmentation performance	The study utilized the scans with a medium Field Of View (FOV) which is 3D sub-volume from CBCT scans, however, no mechanism for localization of medium FOV is provided.