The Saudi Dental Journal 36 (2024) 404–412

Contents lists available at ScienceDirect

The Saudi Dental Journal

journal homepage: www.ksu.edu.sa
www.sciencedirect.com

Review Article

The diagnostic performance of impacted third molars in the mandible: A

review of deep learning on panoramic radiographs
Amalia Nur Faadiya a, Rini Widyaningrum b, *, Pingky Krisna Arindra c, Silviana Farrah Diba b
a
Dental Medicine Study Program, Faculty of Dentistry, Universitas Gadjah Mada, Yogyakarta, Indonesia
b
Department of Dentomaxillofacial Radiology, Faculty of Dentistry, Universitas Gadjah Mada, Yogyakarta, Indonesia
c
Department of Oral and Maxillofacial Surgery, Faculty of Dentistry, Universitas Gadjah Mada, Yogyakarta, Indonesia

A R T I C L E I N F O A B S T R A C T

Keywords: Background: Mandibular third molar is prone to impaction, resulting in its inability to erupt into the oral cavity.
Mandibular canal The radiographic examination is required to support the odontectomy of impacted teeth. The use of computer-
Radiograph aided diagnosis based on deep learning is emerging in the field of medical and dentistry with the advancement of
Panoramic
artificial intelligence (AI) technology. This review describes the performance and prospects of deep learning for
Deep learning
the detection, classification, and evaluation of third molar-mandibular canal relationships on panoramic
Third molar
Impacted radiographs.
Methods: This work was conducted using three databases: PubMed, Google Scholar, and Science Direct. Following
the literature selection, 49 articles were reviewed, with the 12 main articles discussed in this review.
Results: Several models of deep learning are currently used for segmentation and classification of third molar
impaction with or without the combination of other techniques. Deep learning has demonstrated significant
diagnostic performance in identifying mandibular impacted third molars (ITM) on panoramic radiographs, with
an accuracy range of 78.91% to 90.23%. Meanwhile, the accuracy of deep learning in determining the rela­
tionship between ITM and the mandibular canal (MC) ranges from 72.32% to 99%.
Conclusion: Deep learning-based AI with high performance for the detection, classification, and evaluation of the
relationship of ITM to the MC using panoramic radiographs has been developed over the past decade. However,
deep learning must be improved using large datasets, and the evaluation of diagnostic performance for deep
learning models should be aligned with medical diagnostic test protocols. Future studies involving collaboration
among oral radiologists, clinicians, and computer scientists are required to identify appropriate AI development
models that are accurate, efficient, and applicable to clinical services.

1. Introduction cysts and tumors, and traumatic occlusion (Borle, 2014).

Odontectomy is the most common procedure for treating impacted
Impacted teeth are teeth that are partially or cannot fully erupt due teeth by removes impacted teeth or the remaining roots that cannot be
to obstruction by bone or soft tissue or both (Alfadil and Almajed, 2020). extracted by conventional treatments (Wayland, 2018). The categori­
The fact that the third molar is the last teeth to errupt often face greater zation of ITM in the mandible holds significance in assessing the
challenges compared to other teeth due to limited space within the complexity of odontectomy procedures. The Winter’s classifications and
mandibular dental arch (Swift and Nelson, 2012). The global prevalence Pell & Gregory’s are the prevailing method for classifying the ITM
of impacted third molars (ITM) in the mandible is up to 73.5 % (Sujon (Borle, 2014). Fig. 1 demonstrated Winter’s classification that catego­
et al., 2022). Impacted third molars can induce for various dental and rizes ITM based on angulation relative to the long axis of mandibular
oral diseases, such as pericoronitis, caries, periodontal pocket formation, second molars. The classification proposed by Pell & Gregory is utilized
anchorage loss, root damage to adjacent teeth, formation of odontogenic to categorize mandibular ITM by considering the distance between the

Peer review under responsibility of King Saud University. Production and hosting by Elsevier
* Corresponding author at: Department of Dentomaxillofacial Radiology, Faculty of Dentistry, Universitas Gadjah Mada, Denta Street, No. 1, Sleman Regency,
Yogyakarta, Indonesia.
E-mail address: [email protected] (R. Widyaningrum).

https://doi.org/10.1016/j.sdentj.2023.11.025
Received 20 June 2023; Received in revised form 21 November 2023; Accepted 23 November 2023
Available online 25 November 2023
1013-9052/© 2023 THE AUTHORS. Published by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

cervical line and occlusal of the second molar and the depth of the ITM, the diagnosis of ITM in the mandible using panoramic radiographs.
as well as the available space for the crown of the ITM in relation to the Despite the limitations of panoramic radiograph as a 2-dimensional
mandibular ramus (Fig. 1) (Borle, 2014). radiography modality, it remains unknown the significance deep
Common practice in dentistry is to utilize panoramic radiography learning is in supporting the treatment of ITM diagnosed using pano­
imaging of impacted mandibular third molars, allowing for compre­ ramic radiographs. This review further explores the advancement of
hensive view of the dental arch and their surrounding structures before technologies using AI and provides information on the application and
determining odontectomy. Using panoramic radiograph, one can iden­ prospects of deep learning for the radiographic assessment for the
tify the alignment of ITM in the mandible with nearby anatomical detection, classification, and evaluation of the positions of ITM in rela­
structures (Nagaraj et al., 2016). tion to MC.
Knowledge and technology of artificial intelligence (AI) has rapidly
developed; particularly in, assisting in fulfilling human needs, com­ 2. Materials and methods
puters and technology are utilized to imitate human intelligent action
and analytical thinking (Amisha et al., 2019). In the medical field, there PubMed, Science Direct, and Google Scholar databases were used to
has been a significant advancement in the utilization of deep learning- conduct literature searching in this narrative review. Literature search
based computer-aided diagnosis (CAD) systems, which are employed was performed using the keywords, “third molars,” “mandibular nerve,”
for object detection, image analysis, and data processing. These systems “impacted tooth,” “dental radiography,” “artificial intelligence,” and
enable direct data analysis without the need of expert (Yasa et al., 2021). “deep learning”. Manual search was also used for literature search. The
Recently, deep learning has been investigated to provide a second inclusion criteria were literatures published in English between 2012
opinion for the diagnosis of pneumonia based on X-ray images (Mandeel and 2023. The exclusion criteria included articles that were not fully
et al., 2022) and improve the diagnostic performance for skin lesion accessible, excluded research methods, and contained only abstracts.
classification using dermoscopy images (Pratiwi et al., 2021). Con­ In the context of deep learning for computer vision, detection refers
volutional neural networks (CNNs) are subsets of deep learning which to the task of identifying and localizing objects of interest within images.
emulate human intelligence in learning and problem solving, allowing It involves not only recognizing what objects are present in the input
dentists and radiologists to interpret radiographs more efficiently and data but also determining their precise locations in terms of bounding
accurately for caries detection and classification (Vinayahalingam et al., boxes. In this review, deep learning detects the incidence of mandibular
2021b, 2021a), perform teeth detection and numbering (Estai et al., ITM on panoramic radiographs. In deep learning, classification is the
2022), predict third molar eruption (Vranckx et al., 2020), predict time process of categorizing input data into predefined classes. The primary
to extraction of third molar (Kwon et al., 2022), detect apical lesion goal of classification is to assign labels to input instances based on their
(Ekert et al., 2019), execute segmentation on panoramic radiograph for features or characteristics. In this review, deep learning relies on the Pell
periodontitis (Widyaningrum et al., 2022) and detect the periodontal & Gregory and Winter classifications as the fundamental basis for clas­
disease (Lee et al., 2018), as well as classify maxillary sinus (Murata sifying ITM in the mandible. Meanwhile, evaluation refers to the task of
et al., 2019). Deep learning has also studied in several studies to classify categorizing the relationships between ITM and MC, such as whether
mandibular third molars and analyze the correlation between ITM and they are in contact or not.
mandibular canal (MC) automatically with datasets in the form of
panoramic radiograph images using several architectures (Celik, 2022; 3. Results
Choi et al., 2022; Fukuda et al., 2020; Sukegawa et al., 2022a; Yoo et al.,
2021; Zhu et al., 2021). This review was included 49 articles that were elected based on the
The societal desire for high-quality healthcare has driven techno­ predetermined exclusion and inclusion criteria. The present discussion
logical advancements in the field of health. This technology supports focuses on the 12 main articles (Fig. 2). The main articles were original
more accurate and faster diagnoses also ensures that patients can receive articles reporting studies of detection, classification, and the evaluation
the best health care with minimum risk. One significant development in of the association between mandibular third molars and MC.
pursuit of this goal is the utilization of deep learning in the medical field. Several studies of deep learning for the detection, classification, and
Deep learning technology enables the processing of complex data and the evaluation of ITM in the MC have been conducted over the past five
deeper analysis in disease diagnosis. By utilizing deep learning algo­ years. Since 2019, the number of articles has increased annually, and
rithms, dentists and medical professionals can make more precise and these publications culminated in 2022 with seven articles on the
quicker decisions, ultimately enhancing patient care. detection, classification, and evaluation of ITM to the MC by deep
In the past decade, deep learning has been investigated for improving learning. Fig. 3 illustrates the distribution of studies on deep learning for

Fig. 1. Winter’s classification and Pell & Gregory’s classifications of ITM in the mandible.

405
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

Fig. 2. Literature searching and selection.

Fig. 3. Distribution of articles on deep learning applications for the detection, classification, and evaluation of ITM on panoramic radiograph during the last five
years (2019 to January 2023).

panoramic radiographic evaluation of impacted mandibular third mo­ the detection, classification, and evaluation of ITM. Based on the results
lars over the last five years (2019 to January 2023). of a literature search, only one article discussed the detection (seg­
mentation of panoramic radiographs) of mandibular ITM. Thus, we
4. Discussion combined the discussion in this review regarding detection with the
classification of ITM.
We subdivided the articles based on the use of deep learning on
panoramic radiographs in accordance with the study objective, that is,

406
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

4.1. Deep learning performance on panoramic radiographs for detection intersection over union and represents the accuracy prediction of the
and classification of mandibular ITM bounding box. It is determined as the percentage of overlap between the
ground truth area and the estimated bounding box area (Celik, 2022).
Deep learning is widely used for medical image segmentation to Segmentation can be performed manually, but the automation of image
assist the diagnosis of several diseases. Image segmentation is needed to segmentation can reduce the effort and time required to complete the
detect and separate objects of interest in an image (Ronneberger et al., diagnosis process of oral diseases through radiographic images.
2015). Deep learning can detect and classify impacted teeth, such as Previous studies used deep learning to develop an automatic
canine (Aljabri et al., 2022), mesiodens (Jeon et al., 2022), supernu­ mandibular third molar impaction classification system with several
merary in the maxilla (Kuwada et al., 2020), and dental anomalies architectures, such as ResNet-34 (Yoo et al., 2021), VGG-16 (Maruta
(Okazaki et al., 2022). Based on the results of the literature search, one et al., 2023; Sukegawa et al., 2022a), and YOLOv3 (Celik, 2022)
article discussed deep learning using the U-Net model for ITM detection (Table 1). Based on the review results in Table 1, deep learning shows a
in the mandible, which showed a high performance with an average dice high performance in the classification of mandibular ITM (>78.91 %)
coefficient score of 93.6 % and a Jaccard index of 88.1 % (Vinayaha­ (Yoo et al., 2021). The use of deep learning resulted in an accuracy of >
lingam et al., 2019). The dice coefficient was used to measure the level 86 % in Winter’s classification, >82.03 % in the Pell & Gregory’s clas­
of similarity between the results of the detection of ITM in the mandible, sification based on the remaining space, and > 78.91 % in the classifi­
which was performed manually by experts (as ground truth), and the cation based on position of mandibular ITM (Sukegawa et al., 2022a;
findings of automatic segmentation from deep learning algorithms Yoo et al., 2021).
(Widyaningrum et al., 2022). The Jaccard index is also known as the As shown in Table 1, a study on the mandibular ITM with Winter’s

Table 1
Diagnostic Performance of Deep Learning for Detection and Classification of Mandibular ITM on Panoramic Radiograph.
Authors (Year) Detection / Classification Deep Learning Models Number of datasets Diagnostic Performance

Accuracy Others

Vinayahalingam et al. (2019) Detection of impacted third molar U-Net 81 Dice coefficient: 93.6
Jaccard index: 88.1
Sensitivity: 94.7
Specificity: 99.9
Yoo et al. (2021) Winter (Mesioangular) ResNet-34 600 90.23 Sensitivity: 94.15
Specificity: 92.67
(Horizontal) Sensitivity: 89.53
Specificity: 97.65
(Vertikal) Sensitivity: 94.84
Specificity: 95.24
(Distoangular)* –
Pell & Gregory (Class) 82.03
I Sensitivity: 71.69
Specificity: 94.22
II Sensitivity: 90.37
Specificity: 69.52
III Sensitivity: 61.36
Specificity: 98.29
Pell & Gregory (Position) 78.91
A Sensitivity: 88.13
Specificity: 92.05
B Sensitivity: 72.77
Specificity: 84.12
C Sensitivity: 78.63
Specificity: 90.89
Celik, (2022) Winter (Mesioangular) YOLOv3 440 86 Precision: 84.9–90.8
AP: 96–98.4
Recall: 95–98.7
(Horizontal) Precision: 88–96
AP: 98–99.5
Recall: 83.3–100
Sukegawa et al., (2022a) Winter VGG-16 1330 86.63 Precision: 85.59
Recall: 80.03
F1-score: 81.38
AUC: 98.01
Pell & Gregory (Class) 85.41 Precision: 85.88
Recall: 85.44
F1-score: 85.38
AUC: 96.38
(Position) 88.95 Precision: 88.24
Recall: 88.77
F1-score: 88.31
AUC: 97.39
Maruta et al. (2023) Winter VGG-16 1180 79.59 F1-scrore: 64.23
AUC: 95.49
Pell & Gregory (Class) 86.09 F1-scrore: 76.24
AUC: 93.34
(Position) 84.32 F1-scrore: 81.56
AUC: 93.95

* Reveals that the existing data were insufficient in that circumstance; AP: Average precision; AUC: Area under the Receiver Operating Characteristic (ROC) curve.

407
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

and Pell & Gregory’s classifications attained the highest accuracy of specificity using positive predictive value (PPV) calculations. Further­
90.23 % when using the ResNet-34 architecture but used fewer datasets more, the studies reviewed in this work did not use a consistent math­
compared with other research (Yoo et al., 2021). Previous study ematical formula (Ariji et al., 2022; Buyuk et al., 2022; Celik, 2022; Choi
(Sukegawa et al., 2022a) used a single-stage technique with YOLOv3 et al., 2022; Maruta et al., 2023; Takebe et al., 2022; Vinayahalingam
compared with two-stage technique with Faster R-CNN and the addition et al., 2019; Yoo et al., 2021; Zhu et al., 2021). Celik (2022), Choi et al.
of three backbone architectures, namely ResNet50, AlexNet, and VGG16 (2022), Takebe et al., 2022, and Zhu et al. (2021), used PPV calculations
(Celik, 2022). The accuracy of the single-stage technique outperformed to determine precision in their research. A sensitivity calculation is used
that of the two-stage technique by 86 %. This study demonstrated that by Zhu et al. (2021), Takebe et al. (2022) and Choi et al. (2022).
YOLOv3 is effective in classifying ITM based on Winter’s classification. Meanwhile, Celik (2022) used the true positive formula divided by all
The use of data augmentation is one of the ideas proposed to increase the the data as a recall. Buyuk et al. (2022), Vinayahalingam et al. (2021a)
accuracy of deep learning. Flip, rotation, and adjustment of the bright­ provided calculations in accordance with the guidelines for medical
ness, sharpness, contrast and additional images are performed using diagnostic tests in equations (1), (2), (3), (4), and (5) in Table 2 (Lee­
paint software to mimic dental fillings or caries (Maruta et al., 2023). flang, 2014).
The performance indicator of deep learning for the classification In the discipline of informatic and computer science, the confusion
ITM, such as sensitivity, specificity, precision, recall, and accuracy, matrix is a common standard used to assess the deep learning perfor­
cannot be compared with one another because they use the different mance. This method is similar to the diagnostic test methods used in
classification indicators (Table 2) (Celik, 2022; Sukegawa et al., 2022a; medical and health studies. Both confusion matrix and diagnostic test
Yoo et al., 2021). Sukegawa et al., (2022a) and Maruta et al. (2023) used methods are concerned with specificity, sensitivity, negative predictive
the same classification indicator and deep learning architecture, thus, value (NPV), positive predictive value (PPV), and accuracy. The
these studies can be compared. Using larger datasets, Sukegawa et al., confusion matrix includes other terms, such as recall/true positive rate
(2022a) reported accuracy better than Maruta et al. (2023). (which is the same as sensitivity), true negative rate (which is the same
Research was continually developed to compare a single-task model as specificity), and precision (which is the same as PPV) (Chicco et al.,
that classifies impacted mandibular third molars one by one with the 2021). In the field of computer science, sensitivity (true positive rate)
Multi 3Task model, which is a deep learning model that investigates means how well the deep learning algorithm can identify images in their
multiple classifications, for simultaneous several predictions of diag­ category. Specificity is the capability of the deep learning algorithm to
nosis (Sukegawa et al., 2022a). Multi 3Task classifies ITM based on accurately detect images outside of their class. The PPV parameter in­
Winter’s classification and Pell & Gregory’s class and position simulta­ dicates the percentage of actually positive data among data that the
neously. By performing several tasks simultaneously, the multitask algorithm deep learning considers to be positive. The NPV denotes the
model can reduce computation costs; however, similar to previous ratio of actual negative data to the data that deep learning algorithms
research conducted by Celik (2022), the use of the multitask model did interpret as negative. Accuracy is used to calculate the overall deep
not exceed the performance of the single-task model. A significant drop learning performance (Buyuk et al., 2022; Vinayahalingam et al.,
in performance occurred, particularly in Multi 3Task. Winter’s and Pell 2021a). Referring to Table 2, the calculation method is the same from a
& Gregory’s classifications have different classification characteristics. computational and medical perspective but only for the calculation of
Thus, differences in the areas of interest resulted led to a decrease in accuracy.
performance of the multitask model in comparison to the single-task
model. Deep learning was trained to determine Winter’s classification, 4.2. Diagnostic performance of deep learning for evaluating the
which assesses the inclination and angulation of the ITM with feature relationship between ITM and the MC on panoramic radiographs
extraction on all mandibular third molars; Winter’s classification also
has characteristics different from those of Pell & Gregory’s classification Panoramic radiographs are typically used based on the treatment
and uses the area under the mandibular third molars for classification plan for odontectomy to lessen the chance of alveolar nerve impairment
(Sukegawa et al., 2022a). by assessing the association between ITM and MC (Liu et al., 2015).
Nine of the twelve main articles reviewed in this study (Table 2) Technology advancements have enabled the integration of AI into CAD
mentioned the formulas utilized to determine metrics. Some of them systems, which can be used to assess the relationship between the third
used formulas for estimating sensitivity, specificity, NPV, and PPV are molars on either side (bucco-lingual) of the mandible and the mandib­
different from those prescribed by medical diagnostic test calculation ular canal. GoogLeNet, VGG-16, AlexNet (Buyuk et al., 2022; Fukuda
guidelines (Table 2). Yoo et al. (2021) determined the sensitivity and et al., 2020), YOLOv3 (Takebe et al., 2022), YOLOv4 (Zhu et al., 2021),

Table 2
Comparison of the medical diagnostic test and equations for evaluating deep learning performance.
Authors (Year) Sensitivity Specificity PPV NPV Accuracy Other metrics

Formula (Leeflang, 2014) TP TN TP TN TP + TN –

TP + FN TN + FP TP + FP FN + TN TP + FP + TN + FN
Confusion matrix (Chicco et al., Sensitivity = Recall/ True Specificity = True PPV = NPV Accuracy –
2021) Positive Rate Negative Rate Precision
Yoo et al. (2021) – – Sensitivity Specificity Accuracy –
Zhu et al. (2021) Recall – Precision – – F1-score = 2 × precision × recall
/ (precision + recall)
Celik (2022) Recall = TP / All data – Precision – Accuracy –
Choi et al. (2022) Recall – Precision – Accuracy F1-score = 2 × precision × recall
/ (precision + recall)
Buyuk et al. (2022), Sensitivity Specificity PPV NPV Accuracy –
Vinayahalingam et al. (2021a)
Ariji et al. (2022) Sensitivity – – – – –
Takebe et al. (2022) Recall (Sensitivity) – Precision – Accuracy F1-score = 2 × precision × recall
/ (precision + recall)
Maruta et al. (2023) – – – – – F1-score = 2 × precision × recall
/ (precision + recall)

NPV: Negative predictive value; PPV: Positive predictive value; FN: False negative; FP: False positive; TN: True negative; TP: True Positive.

408
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

ResNet-50 (Choi et al., 2022; Sukegawa et al., 2022b) are some deep that are nearly identical to those used in dentistry by classifying pano­
learning architectures that have been investigated for radiographic ramic radiographs of ITM into four categories (Table 3). The fused deep
evaluation of impacted third molar related to the mandibular canal. In learning method was utilized, segmentation was performed using the U-
addition, U-Net is also used for image segmentation of the mandibular Net architecture, and the AlexNet architecture was used to assess the
canal (Ariji et al., 2022; Buyuk et al., 2022; Vinayahalingam et al., relationship between the ITM and MC (Buyuk et al., 2022). This study
2019). also compared the performance of fused deep learning with dentists,
Fukuda et al. (2020), Zhu et al. (2021), Choi et al. (2022), and revealing nearly the same performance, and showed that fused deep
Sukegawa et al., (2022b) evaluated the association between ITM and learning can be used to help in diagnosis.
MC, which was categorized into two distinct types: contact and non- Ariji et al. (2022) use the U-Net and transfer learning technique.
contact. In the study by Fukuda et al. (2020), 600 radiographs with Transfer learning techniques are performed using the learning processes
the GoogLeNet architecture showed a better performance on all pa­ in deep learning models on large datasets (source models) and then
rameters, such as sensitivity (88 ± 6 %), specificity (96 ± 44 %), and transferring knowledge to conduct learning on other deep learning
accuracy (92 ± 5 %), compared with those in the studies of Zhu et al. models with smaller datasets (target models) for the same task (Alzu­
(2021) and Choi et al. (2022). High sensitivity means that deep learning baidi et al., 2021). This study conducted by Ariji et al. (2022) shows the
can correctly predict positive results if the radiograph shows contact target model’s deep learning performance is on par with the source
between ITM and MC. High specificity indicates the capability of deep model, especially in 200 datasets indicate that the transfer learning
learning to correctly predict negative results if the impacted the ITM and technique can be a method for transferring knowledge in a relatively
MC is non-contact. The quantity of radiographs as dataset supports the small dataset. Transfer learning techniques can be a solution for devel­
outcomes of this high performance of deep learning because it will oping deep learning using large datasets obtained from several in­
enhance the capacity to learn. stitutions while still protecting the security of patient medical record
Additionally, Fukuda et al. (2020) used a data augmentation process data (Ariji et al., 2022). The study’s findings demonstrate that the target
as part of the process for creating the deep learning algorithms, partic­ model’s deep learning performance is comparable to that of the source
ularly in the case of a restricted quantity of datasets. The use of data model, especially in small datasets. While maintaining the integrity of
augmentation, which is commonly used to improve performance, had an patient medical record data, transfer learning techniques can be used to
impact on accuracy (Maruta et al., 2023). Fukuda et al. (2020) compared construct deep learning using massive datasets collected from many
three deep learning architectures—GooLeNet, AlexNet, and VGG-16—to institutions (Ariji et al., 2022).
evaluate the ITM in relation to the MC. Each architecture was designed
using distinct image from cropped panoramic radiograph pictures 4.3. Prospects for deep learning applications in the Detection,
(Fukuda et al., 2020). Deep learning performs more accurately in terms Classification, and evaluation of mandibular ITM on panoramic
of prediction the small-sized image (England and Cheng, 2019). Datasets radiographs
with larged-size images typically contain more complicated informa­
tion, in this case, additional anatomical images from panoramic radio­ Deep learning has developed because of today’s increasingly
graphs that are not necessary for the learning process (Fukuda et al., powerful hardware and software improvements, and it has been used in
2020). many fields of studies, including oral radiology (Corbella et al., 2021).
Zhu et al. (2021) compared the accuracy of deep learning with CBCT is widely regarded as the preferred method as well as the gold
dentists (Zhu et al., 2021) In addition, deep learning has been used by standard for supporting examinations of ITM. CBCT involves maxillo­
dentists to evaluate the relationship between the ITM and the MC. With facial imaging that uses the most advanced technology; it can display
a precision of 93.88 % and a recall of 92 %, dentists applied deep three-dimensional images from a cross-sectional view to evaluate the
learning to produce the highest performance results. This finding in­ incidence or nonexistence of corticalization between the ITM and MC
dicates that dentists can use deep learning to assess how mandibular (Nasser et al., 2018). Deep learning can also identify and categorize ITM
third molars and MC interact to provide more accurate findings; how­ with the use of CBCT (Borgonovo et al., 2017; Liu et al., 2022; Orhan
ever, further research is required because of the limited use of radio­ et al., 2021). On the other hand, CBCT is expensive and requires larger
graphs and dentists (Zhu et al., 2021). doses of radiation, thus, panoramic radiography remains a widely used
Sukegawa et al., (2022b) analyzed the effect of optimizer algorithms an alternative in the examination of impacted teeth (Whaites and Drage,
with gradient methods, such as sharpness aware minimization (SAM) 2014), in addition as a preliminary step to determining whether CBCT
and stochastic gradient descent (SGD). ResNet 50v2 with an SAM opti­ imaging is necessary. In diagnosis, deep learning may be used in
mizer showed the best performance in this research. Overlapping and conjunction with CBCT (Takebe et al., 2022). However, deep learning
poor generalization performance results were obtained with the use of improves the visualization of the pseudo contact of ITM with the MC on
SGD; meanwhile, SAM enhanced robustness against noise and general­ panoramic radiographs. Therefore, the utilization of CBCT can be
ization performance (Sukegawa et al., 2022b) and Choi et al. (2022) also diminished. (Zhu et al., 2021).
performed similar studies, but with six and five oral surgeons, respec­ AI based on deep learning can be applied in clinical practice to assist
tively. Compared with oral surgeons, ResNet-50 and YOLOv3 performed dentists in making a diagnosis quickly and accurately, minimize mis­
better. interpretations of diagnoses due to high workloads, and speed up the
The low diagnostic performance of oral surgeons is affected by the interpretation process (Zhu et al., 2021). AI will greatly facilitate the
difficulty of using panoramic radiographs to determine the bucco/ screening process in regions with a shortage of radiologists (Celik,
lingual position and vertical relationship of the mandibular third molars 2022). In the future, the detection, classification, and evaluation of ITM
to the mandibular canal (Choi et al., 2022). In the evaluation of the in the mandible using panoramic radiographs can be integrated in recent
relationship between the ITM, radiographic features, such as canal applications to support the diagnosis in dental practice, as well as to
narrowing, dark root apexes, bifid root apexes, root narrowing, inter­ increase the diagnostic accuracy of radiographic examinations. As the
ruption of the mandibular canal wall, canal deflection, and root process of making a diagnosis is still handled by the dentist in charge of
deflection, can be observed (Nasser et al., 2018; Whaites and Drage, the patient, dentists and oral radiologists are still in charge of patient
2014). The position of impacted mandibular third molars that are buccal care. Deep learning or any other AI cannot fully replace dentists or ra­
or lingual to the mandibular canal but not in contact with it can be diologists because of the probability of misdiagnosis, that is, either un­
revealed by the superimposed radiographic appearance (Choi et al., derdiagnosis or overdiagnosis (Roosanty et al., 2022).
2022). Buyuk et al. (2022) addressed the needs of dentists because deep The findings presented in Table 2 demonstrated the inconsistent
learning can assess the association between ITM and MC using standards performance of deep learning in facilitating the diagnosis of impacted

409
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

Table 3
Diagnostic Performance of Deep Learning in Evaluating the Relationship Between ITM and MC.
Authors (Year) Evaluation Relationship with Mandibular Canal Deep Learning Models Numbers of Diagnostic Performance
Dataset
Accuracy Others

Vinayahalingam et al., Segmentation of mandibular canals U-Net 81 – Dice coefficient: 80.5

(2019) Jaccard index: 68.7
Sensitivity: 84.7
Specificity: 96.7
Fukuda et al., (2020) Contact/ GoogLe Net 600 92 ± 5 Sensitivity: 88 ± 6
Superimposed and Specificity: 96 ± 4
Non-contact/ AUC: 93
Non-superimposed
Zhu et al. (2021) Contact and Non-contact YOLOv4 503 – Precision: 88.71
Recall: 91.67
AP: 83.02
F1-score: 90.16
YOLOv4 and Dentist Precision: 93.88
Recall: 92.00
AP: 88.06
F1-score: 92.93
Choi et al. (2022) Contact and Non-contact ResNet-50 571 72.32 Sensitivity: 84.62
Specificity: 55.32
Buccolingual position 80.65 Sensitivity: 86.67
Specificity: 75
Buyuk et al. (2022) Segmentation mandibular canals U-Net 1880 99 Dice coefficient: 91
Jaccard index: 98
Non-contact AlexNet 80 Sensitivity: 74
Specificity: 92
PPV: 79
NPV: 90
Contact Sensitivity: 83
Specificity: 95
PPV: 88
NPV: 94
Superimposed Sensitivity: 86
Specifisity: 88
PPV: 80
NPV: 93
Mandibular third molar roots that crossed the Sensitivity: 67
mandibular canal. Specifisity: 96
PPV: 68
NPV: 96
Takebe et al. (2022) Contact YOLOv3 579 89.4 Precision: 89.1
Recall: 92.5
F1-score: 90.8
Sukegawa et al., (2022b) Contact and Non-contact ResNet50 with SAM 1279 85.5 Precision: 81
Optimizer Recall: 78.5
F1-score: 79.4
AUC: 88.3
ResNet50 with SGD 85 Precision: 80.4
Optimizer Recall: 78.1
F1-score: 78.9
AUC:87.5
ResNet50v2 with SAM 86 Precision: 81.6
Optimizer Recall: 79.1
F1-score: 80
AUC: 89
ResNet50v2 with SGD 85.3 Precision: 80.9
Optimizer Recall: 78.2
F1-score: 79.2
AUC: 88.4
Ariji et al. (2022) Segmentation mandibular canal U-Net with Transfer 881 – Dice coefficient:
Learning 74.9–85.7
Jaccard index:
60.7–79.1
Sensitivity: 71.4–87.1

AP: Average precision; AUC: Area under the ROC curve; NPV: Negative predictive value; PPV: Positive predictive value.

teeth across different studies. Regarding the results presented in Table 2, facilitate the development of AI studies that are tailored to satisfy the
further studies on deep learning for medical and dental applications demands of medical and dental practices. However, dentists and dental
should consider using diagnostic performance calculations that are more specialists participating in deep learning research are still very few (Choi
consistent with the diagnostic test commonly employed in medical and et al., 2022; Zhu et al., 2021). To produce high performance deep
health studies. learning-based CAD that can be applied clinically in dental practice,
Bringing together the fields of computer science, radiology, oral computing experts and dentists should collaborate more in the future to
surgery, and statistical science through collaborative efforts will develop deep learning. In further research, deep learning can be used to

410
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

estimate the difficulty level of ITM extraction and radiographs can be References
used to evaluate the depth of impaction of mandibular ITM. The depth of
mandibular third molars is assessed using two methods: Winters line and Alfadil, L., Almajed, E., 2020. Prevalence of impacted third molars and the reason for
extraction in saudi arabia. Saudi Dent. J. 32, 262–268. https://doi.org/10.1016/j.
second molar guides (Whaites and Drage, 2014). Moreover, third mo­ sdentj.2020.01.002.
lars, which are still partly erupted and far from the mandibular canal, Aljabri, M., Aljameel, S.S., Min-Allah, N., Alhuthayfi, J., Alghamdi, L., Alduhailan, N.,
minimize the risk of complications in odontectomy. Further research on Alfehaid, R., Alqarawi, R., Alhareky, M., Shahin, S.Y., Al Turki, W., 2022. Canine
impaction classification from panoramic dental radiographic images using deep
deep learning to predict the complete eruption position of mandibular learning models. Inform. Med. Unlocked 30, 100918. https://doi.org/10.1016/j.
ITM that are still in dentition will be beneficial in clinical practice (Zhu imu.2022.100918.
et al., 2021). Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O.,
Santamaría, J., Fadhel, M.A., Al-Amidie, M., Farhan, L., 2021. Review of deep
Deep learning can enhance assistance for clinicians by incorporating learning: concepts, cnn architectures, challenges, applications, Future Directions.
predictive factors such as the complexity of wisdom tooth extraction J. Big Data 8, 1–74. https://doi.org/10.1186/s40537-021-00444-8.
(including the potential need for root division), estimating extraction Amisha, M.P., Pathania, M., Rathaur, V., 2019. Overview of artificial intelligence in
medicine. J. Family Med. Prim. Care 8, 2328–2331. https://doi.org/10.4103/jfmpc.
time, assessing the likelihood of root adhesion, and considering their
jfmpc_440_19.
relationship with the mandibular canal. Nevertheless, acquiring this Ariji, Y., Mori, M., Fukuda, M., Katsumata, A., Ariji, E., 2022. automatic visualization of
data presents challenges as it relies on clinical information. Considering the mandibular canal in relation to an impacted mandibular third molar on
this issue, the development of AI for CAD necessitates multidisciplinary panoramic radiographs using deep learning segmentation and transfer learning
techniques. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 134, 749–757. https://
collaboration among clinicians, radiologists, and computer scientists doi.org/10.1016/j.oooo.2022.05.014.
(Montagnon et al., 2020). From a clinical standpoint, radiologists should Borgonovo, A.E., Rigaldo, F., Maiorana, C., Grossi, G.B., Augusti, D., Re, D., 2017. CBCT
collaborate with oral surgeons to determine the problem formulation, evaluation of the tridimensional relationship between impacted lower third molar
and the inferior alveolar nerve position. Minerva Stomatol. 66, 9–19. https://doi.
which demands precise diagnosis and treatment planning for patients. org/10.23736/S0926-4970.16.03976-X.
Meanwhile, oral radiologists and computer scientists must collaborate Borle, R.M., 2014. Textbook of Oral and Maxillofacial Surgery. Jaypee Brothers Medical
for the development of imaging examinations with the best visualiza­ Publisher (P), New Delhi.
Buyuk, C., Akkaya, N., Arsan, B., Unsal, G., Aksoy, S., Orhan, K., 2022. A fused deep
tion, beginning with annotation, segmentation, training, testing, and learning architecture for the detection of the relationship between the mandibular
validation, ensuring that deep learning works precisely based on the third molar and the mandibular canal. Diagnostics 12, 2018. https://doi.org/
needs of clinicians in dental practice. 10.3390/diagnostics12082018.
Celik, M.E., 2022. Deep learning based detection tool for impacted mandibular third
molar teeth. Diagnostics 12, 942. https://doi.org/10.3390/diagnostics12040942.
5. Conclusion Chicco, D., Tötsch, N., Jurman, G., 2021. The matthews correlation coefficient (mcc) is
more reliable than balanced accuracy, bookmaker informedness, and markedness in
two-class confusion matrix evaluation. BioData Min. 14, 1–22. https://doi.org/
Deep learning-based AI applications show high performance in the
10.1186/s13040-021-00244-z.
detection, classification, and evaluation of mandibular ITM to the MC on Choi, E., Lee, S., Jeong, E., Shin, S., Park, H., Youm, S., Son, Y., Pang, K.M., 2022.
panoramic radiographs. With or without the combined usage of other Artificial intelligence in positioning between mandibular third molar and inferior
techniques, deep learning architectures used are continually being alveolar nerve on panoramic radiography. Sci. Rep. 12, 2456. https://doi.org/
10.1038/s41598-022-06483-2.
developed. Deep learning-based AI can be improved and used in clinical Corbella, S., Srinivas, S., Cabitza, F., 2021. Applications of deep learning in dentistry.
practice to aid dentists in diagnosing patients more accurately. A Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 132, 225–238. https://doi.org/
multidisciplinary approach involving clinicians, radiologists, and com­ 10.1016/j.oooo.2020.11.003.
Ekert, T., Krois, J., Meinhold, L., Elhennawy, K., Emara, R., Golla, T., Schwendicke, F.,
puter scientists must be developed to accomplish this goal. 2019. Deep learning for the radiographic detection of apical lesions. J. Endod. 45,
917–922.e5. https://doi.org/10.1016/j.joen.2019.03.016.
Ethical statement England, J.R., Cheng, P.M., 2019. Artificial intelligence for medical image analysis: a
guide for authors and reviewers. Am. J. Roentgenol. https://doi.org/10.2214/
AJR.18.20490.
No ethical issues were raised during the study presentation. Estai, M., Tennant, M., Gebauer, D., Brostek, A., Vignarajan, J., Mehdizadeh, M., Saha, S.,
2022. Deep learning for automated detection and numbering of permanent teeth on
panoramic images. Dentomaxillofacial Radiol. 51, 20210296. https://doi.org/
CRediT authorship contribution statement 10.1259/dmfr.20210296.
Fukuda, M., Ariji, Y., Kise, Y., Nozawa, M., Kuwada, C., Funakoshi, T., Muramatsu, C.,
Amalia Nur Faadiya: Conceptualization, Methodology, Validation, Fujita, H., Katsumata, A., Ariji, E., 2020. Comparison of 3 deep learning neural
networks for classifying the relationship between the mandibular third molar and
Formal analysis, Investigation, Data curation, Writing – original draft,
the mandibular canal on panoramic radiographs. Oral Surg. Oral Med. Oral Pathol.
Visualization. Rini Widyaningrum: Methodology, Formal analysis, Oral Radiol. 130, 336–343. https://doi.org/10.1016/j.oooo.2020.04.005.
Investigation, Data curation, Writing – original draft. Pingky Krisna Jeon, K.J., Ha, E.G., Choi, H., Lee, C., Han, S.S., 2022. Performance comparison of three
Arindra: Writing – review & editing, Supervision. Silviana Farrah deep learning models for impacted mesiodens detection on periapical radiographs.
Sci. Rep. 12, 1–8. https://doi.org/10.1038/s41598-022-19753-w.
Diba: Writing – review & editing, Validation. Kuwada, C., Ariji, Y., Fukuda, M., Kise, Y., Fujita, H., Katsumata, A., Ariji, E., 2020. Deep
learning systems for detecting and classifying the presence of impacted
Declaration of competing interest supernumerary teeth in the maxillary incisor region on panoramic radiographs. Oral
Surg. Oral Med. Oral Pathol. Oral Radiol. 130, 464–469. https://doi.org/10.1016/j.
oooo.2020.04.813.
The authors declare that they have no known competing financial Kwon, D., Ahn, J., Kim, C.S., Kang, D.O., Paeng, J.Y., 2022. A deep learning model based
interests or personal relationships that could have appeared to influence on concatenation approach to predict the time to extract a mandibular third molar
tooth. BMC Oral Health 22, 571. https://doi.org/10.1186/s12903-022-02614-3.
the work reported in this paper. Lee, J.H., Kim, D.H., Jeong, S.N., Choi, S.H., 2018. Diagnosis and prediction of
periodontally compromised teeth using a deep learning-based convolutional neural
Acknowledgment network algorithm. J. Periodontal. Implant Sci. 48, 114–123. https://doi.org/
10.5051/jpis.2018.48.2.114.
Leeflang, M.M.G., 2014. Systematic reviews and meta-analyses of diagnostic test
This review is part of the first author’s thesis for undergraduate final accuracy. Clin. Microbiol. Infect. 20, 105–113. https://doi.org/10.1111/1469-
project, under supervision of other authors. The authors acknowledge 0691.12474.
Liu, M.-Q., Xu, Z.-N., Mao, W.-Y., Li, Y., Zhang, X.-H., Bai, H.-L., Ding, P., Fu, K.-Y., 2022.
Universitas Gadjah Mada, Indonesia, for supporting this publication
Deep learning-based evaluation of the relationship between mandibular third molar
under grant of Final Project Recognition Grant Number 5075/UN1.P.II/ and mandibular canal on CBCT. Clin. Oral Invest. 26, 981–991. https://doi.org/
Dit-Lit/PT.01.01/2023. 10.1007/s00784-021-04082-5.
Liu, W., Yin, W., Zhang, R., Li, J., Zheng, Y., 2015. Diagnostic value of panoramic
radiography in predicting inferior alveolar nerve injury after mandibular third molar
extraction: a meta-analysis. Aust. Dent. J. 60, 233–239. https://doi.org/10.1111/
adj.12326.

411
A.N. Faadiya et al. The Saudi Dental Journal 36 (2024) 404–412

Mandeel, T.H., Awad, S.M., Naji, S., 2022. Pneumonia binary classification using multi- learning in deep learning-based positioning classification of mandibular third
scale feature classification network on chest X-ray images. IAES Int. J. Artif. Intell. molars. Sci. Rep. 12, 1–10. https://doi.org/10.1038/s41598-021-04603-y.
(IJ-AI) 11, 1469–1477. https://doi.org/10.11591/ijai.v. Sukegawa, S., Tanaka, F., Hara, T., Yoshii, K., Yamashita, K., Nakano, K., Takabatake, K.,
Maruta, N., Morita, K.I., Harazono, Y., Anzai, E., Akaike, Y., Yamazaki, K., Tonouchi, E., Kawai, H., Nagatsuka, H., Furuki, Y., 2022b. Deep learning model for analyzing the
Yoda, T., 2023. Automatic machine learning-based classification of mandibular third relationship between mandibular third molar and inferior alveolar nerve in
molar impaction status. J. Oral Maxillofac. Surg. Med. Pathol. 35, 327–334. https:// panoramic radiography. Sci. Rep. 12, 16925. https://doi.org/10.1038/s41598-022-
doi.org/10.1016/j.ajoms.2022.12.010. 21408-9.
Montagnon, E., Cerny, M., Cadrin-Chênevert, A., Hamilton, V., Derennes, T., Ilinca, A., Swift, J.Q., Nelson, W.J., 2012. The nature of third molars: are third molars different
Vandenbroucke-Menu, F., Turcotte, S., Kadoury, S., Tang, A., 2020. deep learning than other teeth? Atlas Oral Maxillof. Surg. Clin. 20, 159–162. https://doi.org/
workflow in radiology: a primer. Insights Imaging. https://doi.org/10.1186/s13244- 10.1016/j.cxom.2012.07.003.
019-0832-5. Takebe, K., Imai, T., Kubota, S., Nishimoto, A., Amekawa, S., Uzawa, N., 2022. Deep
Murata, M., Ariji, Y., Ohashi, Y., Kawai, T., Fukuda, M., Funakoshi, T., Kise, Y., learning model for the automated evaluation of contact between the lower third
Nozawa, M., Katsumata, A., Fujita, H., Ariji, E., 2019. Deep-learning classification molar and inferior alveolar nerve on panoramic radiography. J. Dent. Sci. https://
using convolutional neural network for evaluation of maxillary sinusitis on doi.org/10.1016/j.jds.2022.12.008.
panoramic radiography. Oral Radiol. 35, 301–307. https://doi.org/10.1007/ Vinayahalingam, S., Xi, T., Bergé, S., Maal, T., de Jong, G., 2019. Automated detection of
s11282-018-0363-7. third molars and mandibular nerve by deep learning. Sci. Rep. 9, 9007. https://doi.
Nagaraj, T., Keerthi, I., James, L., Shruthi, R., Balraj, L., Bhavana, T.V., 2016. Visibility of org/10.1038/s41598-019-45487-3.
mandibular anatomical landmarks in panoramic radiography: a retrospective study. Vinayahalingam, S., Kempers, S., Limon, L., Deibel, D., Maal, T., Hanisch, M., Bergé, S.,
J. Med. Radiol. Pathol. Surg. 2, 14–17. https://doi.org/10.15713/ins.jmrps.57. Xi, T., 2021a. Classification of caries in third molars on panoramic radiographs using
Nasser, A., Alomar, A., AlOtaibi, N., Altamimi, A., 2018. Correlation of panoramic deep learning. Sci. Rep. 11 https://doi.org/10.1038/s41598-021-92121-2.
radiograph and CBCT findings in assessment of relationship between impacted Vinayahalingam, S., Kempers, S., Limon, L., Maal, T., Bergé, S., Xi, T., Hanisch, M.,
mandibular third molars and mandibular canal in Saudi population. Dent. Oral 2021b. The automatic detection of caries in third molars on panoramic radiographs
Craniofac. Res. 4, 1–5. https://doi.org/10.15761/docr.1000256. using deep learning: a pilot study. Res Sq. https://doi.org/10.21203/rs.3.rs-379636/
Okazaki, S., Mine, Y., Iwamoto, Y., Urabe, S., Mitsuhata, C., Nomura, R., Kakimoto, N., v1.
Murayama, T., 2022. Analysis of the feasibility of using deep learning for multiclass Vranckx, M., Van Gerven, A., Willems, H., Vandemeulebroucke, A., Leite, A.F., Politis, C.,
classification of dental anomalies on panoramic radiographs. Dent. Mater. J. 41, Jacobs, R., 2020. Artificial intelligence (AI)-driven molar angulation measurements
889–895. https://doi.org/10.4012/dmj.2022-098. to predict third molar eruption on panoramic radiographs. Int. J. Environ. Res.
Orhan, K., Bilgir, E., Bayrakdar, I.S., Ezhov, M., Gusarev, M., Shumilov, E., 2021. Public Health 17, 3716. https://doi.org/10.3390/ijerph17103716.
Evaluation of artificial intelligence for detecting impacted third molars on cone- Wayland, J., 2018. Impacted Third Molars. John Wiley & Sons, Pondicherry.
beam computed tomography scans. J. Stomatol. Oral Maxillofac. Surg. 122, Whaites, E., Drage, N., 2014. Essentials of Dental Radiography and Radiology, fifth ed.
333–337. https://doi.org/10.1016/j.jormas.2020.12.006. Elsevier, China.
Pratiwi, R.A., Nurmaini, S., Rini, D.P., Rachmatullah, M.N., Darmawahyuni, A., 2021. Widyaningrum, R., Candradewi, I., Aji, N.R.A.S., Aulianisa, R., 2022. Comparison of
Deep ensemble learning for skin lesions classification with Convolutional Neural Multi-Label U-Net and Mask R-CNN for panoramic radiograph segmentation to
Network. IAES International. J. Artif. Intell. 10, 563–570. https://doi.org/10.11591/ detect periodontitis. Imaging Sci. Dent. 52, 383–391. https://doi.org/10.5624/
ijai.v10.i3.pp563-570. isd.20220105.
Ronneberger, O., Fischer, P., Brox, T. 2015. U-net: Convolutional networks for Yasa, Y., Çelik, Ö., Bayrakdar, I.S., Pekince, A., Orhan, K., Akarsu, S., Atasoy, S.,
biomedical image segmentation. In Medical Image Computing and Computer- Bilgir, E., Odabaş, A., Aslan, A.F., 2021. An artificial intelligence proposal to
Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, automatic teeth detection and numbering in dental bite-wing radiographs. Acta
Germany, October 5-9, 2015. Proceedings, Part III 18 (pp. 234-241). Springer Odontol. Scand. 79, 275–281. https://doi.org/10.1080/00016357.2020.1840624.
International Publishing. Yoo, J.H., Yeom, H.G., Shin, W.S., Yun, J.P., Lee, J.H., Jeong, S.H., Lim, H.J., Lee, J.,
Roosanty, A., Widyaningrum, R., Diba, S.F., 2022. Artificial intelligence based on Kim, B.C., 2021. Deep learning based prediction of extraction difficulty for
convolutional neural network for detecting dental caries on bitewing and periapical mandibular third molars. Sci. Rep. 11, 1954. https://doi.org/10.1038/s41598-021-
radiographs. J. Radiol. Dentomaksilof. Indonesia (JRDI) 6, 89–94. https://doi.org/ 81449-4.
10.32793/jrdi.v6i2.867. Zhu, T., Chen, D., Wu, F., Zhu, F., Zhu, H., 2021. Artificial intelligence model to detect
Sujon, M.K., Alam, M.K., Rahman, S.A., Noor, S.N.F.M., 2022. Third molar impactions real contact relationship between mandibular third molars and inferior alveolar
prevalence and pattern among adults using 5923 digital orthopantomogram. nerve based on panoramic radiographs. Diagnostics 11, 1664. https://doi.org/
Bangladesh J. Med. Sci. 21, 717–729. 10.3390/DIAGNOSTICS11091664.
Sukegawa, S., Matsuyama, T., Tanaka, F., Hara, T., Yoshii, K., Yamashita, K., Nakano, K.,
Takabatake, K., Kawai, H., Nagatsuka, H., Furuki, Y., 2022a. Evaluation of multi-task

412