Received: 3 May 2021 | Revised: 14 June 2021 | Accepted: 4 July 2021
DOI: 10.1111/jop.13227
ORIGINAL ARTICLE
Automatic classification and detection of oral cancer in
photographic images using deep learning algorithms
Kritsasith Warin1 | Wasit Limprasert2 | Siriwan Suebnukarn3 |
Suthin Jinaporntham4 | Patcharapon Jantana5
1
Division of Oral and Maxillofacial
Surgery, Faculty of Dentistry, Thammasat Abstract
University, Pathum Thani, Thailand
Background: Oral cancer is a deadly disease among the most common malignant tu-
2
College of Interdisciplinary Studies,
Thammasat University, Patum Thani,
mors worldwide, and it has become an increasingly important public health problem
Thailand in developing and low-to-middle income countries. This study aims to use the con-
3
Faculty of Dentistry, Thammasat volutional neural network (CNN) deep learning algorithms to develop an automated
University, Pathum Thani, Thailand
4 classification and detection model for oral cancer screening.
Department of Oral and Maxillofacial
Surgery, Faculty of Dentistry, Khon Kaen Methods: The study included 700 clinical oral photographs, collected retrospectively
University, Khon Kaen, Thailand
5
from the oral and maxillofacial center, which were divided into 350 images of oral
StoreMesh, Thailand Science Park,
Pathum Thani, Thailand squamous cell carcinoma and 350 images of normal oral mucosa. The classification
and detection models were created by using DenseNet121 and faster R-CNN, respec-
Correspondence
Kritsasith Warin, FRCDT (Oral and tively. Four hundred and ninety images were randomly selected as training data. In ad-
Maxillofacial Surgery), Division of Oral and dition, 70 and 140 images were assigned as validating and testing data, respectively.
Maxillofacial Surgery, Faculty of Dentistry,
Thammasat University, Pathum Thani, Results: The classification accuracy of DenseNet121 model achieved a precision of
12121, Thailand. 99%, a recall of 100%, an F1 score of 99%, a sensitivity of 98.75%, a specificity of
Email: warin@[Link]
100%, and an area under the receiver operating characteristic curve of 99%. The de-
Funding information tection accuracy of a faster R-CNN model achieved a precision of 76.67%, a recall of
Thammasat University, Grant/Award
Number: TUFT24/2564 82.14%, an F1 score of 79.31%, and an area under the precision-recall curve of 0.79.
Conclusion: The DenseNet121 and faster R-CNN algorithm were proved to offer the
acceptable potential for classification and detection of cancerous lesions in oral pho-
tographic images.
KEYWORDS
artificial intelligence, deep learning, oral cancer, telemedicine
1 | I NTRO D U C TI O N to lifestyle behaviors such as smoking, alcohol consumption, and
betel quid chewing. 2 The management of oral cancer, including
Oral cancer is a deadly disease which is the 17th most com- surgery radiotherapy and/or chemotherapy, especially at an ad-
mon cancer worldwide and the 11th in Asia, with approximately vanced stage, can lead to more morbidities and be very costly. 3 In
380 000 new cases and nearly half of the number of these new addition, the prognosis for oral cancer is poor. The current study
cases of death in 2020.1 Two-thirds of oral cancers have been showed that the overall 5 years survival rate for oral cancer is ap-
found in developing and low-to-middle income countries, espe- proximately 52% but will decrease to 31.2% and 12.5% if there are
cially in Southeast Asia and South Asia which may be attributable regional and distant metastases, respectively.4 Early diagnosis of
© 2021 John Wiley & Sons A/S. Published by John Wiley & Sons Ltd
J Oral Pathol Med. 2021;50:911–918. [Link]/journal/jop | 911
912 | WARIN et al.
oral cancer is very important so that the morbidity and mortality 2 | M ATE R I A L S A N D M E TH O DS
rate of patients can be reduced.
As mentioned earlier, oral cancer occurs mainly in low-
to- 2.1 | Ethical approval
middle income countries where the number of oral cancer special-
ists is limited. 2 Patients in remote areas also have limited access to This study was approved by the ethics review board of our university
appropriate diagnosis and treatment. Although the gold standard and was performed in accordance with the tenets of the Declaration
of oral cancer diagnosis is pathologically proven, 3 an abnormal- of Helsinki. Informed consent was waived because of the retrospec-
ity in clinical appearances may also be a clue for the clinician to tive nature of the fully anonymized images.
screen suspected malignant lesions in the oral cavity. Nowadays,
telemedicine plays an important role in helping oral cancer spe-
cialists communicate with general practitioners and patients in 2.2 | Dataset
remote areas. This remote consultation may improve the accuracy
of oral cancer screening and appropriate referral. 5 It would be of The dataset consisted of 700 clinical oral photographs retrospec-
great benefit to take this concept a step further by integrating an tively collected from our oral and maxillofacial surgery center be-
automated recognition system utilizing artificial intelligence 6 to tween 2009 and 2018. The 700 images were divided into 350 images
analyze oral photographic images. of oral squamous cell carcinoma (OSCC) and another 350 normal oral
Artificial intelligence (AI) is a branch of computer science mucosae. All of the OSCC images were biopsy proven as the gold
which can be defined as the ability for a computer to mimic the standard for oral cancer diagnosis.
cognitive abilities of a human being. AI corresponds to a large
array of techniques. Among them, deep learning is a potential
disruptive technology that attempts to model high-
l evel ab- 2.3 | Reference data
stractions in medical images to determine diagnostic meaning.
Deep learning, specifically as implemented using convolutional Our reference data were clinical oral photographs of OSCC and nor-
neural networks (CNNs), has become a conventional technique mal oral mucosa. The OSCC images that were used for analysis in
for classifying, detecting, and segmenting the objects in med- this study are OSCC stage I-IV according to the TNM clinical staging
7
ical images. A deep learning-b ased image classification is one system as proposed by the American Joint Committee on Cancer
of the useful tools for classifying medical diseases in clinical (AJCC).15 The OSCC images are characterized as ulcerative, exo-
images and X-rays. For example, the DenseNet algorithm has phytic, or endophytic in various areas of the oral cavity,16 including
8
been used to classify mass in breast mammography images. A the lip, upper and lower alveolar ridge, buccal mucosa, tongue, and
regional convolutional neural networks (R-C NN), which is part of hard palate. All OSCC images have been pathologically proven. The
CNNs, have been proven to have the highest accuracy for object images of normal oral mucosa were also selected from the various
detection.9 Faster R-C NN, in particular, works reasonably well area in the oral cavity.
with a small dataset.10 In medicine, Faster R-C NN has been used All photographic images were uploaded to the VisionMarker
for automated detection of the abnormality in X-ray and clin- server. VisionMarker is a private web application for image an-
ical images such as lung nodules in computerized tomography notation. The public version is available on GitHub.17 The lesion
scans (CT-s cans), esophageal adenocarcinoma in high-d efinition boundaries of the OSCC images were annotated by three oral and
white light endoscopy (HD-W LE) images,11 and skin diseases maxillofacial surgeons. Due to the differences in manual segmen-
12
from clinical photography. In dentistry, there are a few stud- tation from one expert to another, we used the largest area of in-
ies on automated detection using Faster R-C NN mainly for the tersection between all of the surgeon's annotations in combination
detection of abnormalities on the radiograph. The faster R-C NN with the pathological result of OSCC as the ground truth in the deep
trained on a limited amount of labeled imaging data performed learning model training, validation, and testing (Figure 1).
satisfactorily in detecting periodontally compromised teeth in
digital panoramic radiographs.13 Another study used CNN mod-
els for detecting and classifying the presence of impacted su- 2.4 | Experiments
pernumerary teeth in the maxillary incisor region on panoramic
radiographs.14 Deep learning, especially convolutional neural network (CNN), has
The purpose of this study is to develop an automated classi- been shown to be effective in object classification and detection. In
fication and detection system for oral cancer screening using the this work, we adopt DenseNet121 to classify OSCC apart from the
DenseNet121 and Faster R-CNN algorithm. The use of this system is normal oral mucosa and used Faster R-CNN to detect OSCC on the
expected to assist clinicians in detecting suspected malignant lesions oral photograph.
in the oral cavity. In combination with telemedicine, remote detec- Faster R-CNN was developed by Ren et al.10 as a deep learning-
tion of oral cancer can help clinicians detect oral cancer at an early based object detection system by combining region proposal net-
stage and improve the accuracy of referral systems. work (RPN) and the previous object detection system, Fast R-CNN,
WARIN et al. | 913
F I G U R E 1 Examples of the oral photographic images from the dataset showing (A) non-cancer image, (B) OSCC image, and (C) the largest
intersection area between the annotations from three different surgeons
into one single network by sharing their convolutional features lead- in this experiment is DenseNet121 with pre-trained weight from
ing to a more real-time method. The proposed RPN generates re- ImageNet. The object detection experiment used the annotated
gion proposals for each location using the last feature map produced image from VisionMarker. The annotated images were identified
from the CNN based on anchor boxes. The anchor boxes are detec- bounding boxes locations of the lesion areas; then, the pairs of image
tion boxes that have different sizes and ratios that are compared to and annotation were ready for the training process. The training was
the ground truth during the training process. performed on an on-premise server with 2 of GPU, TitanXP 12GB,
Dense convolutional network or DenseNet was introduced by Nvidia Driver: 450.102 and CUDA: 11.0. The neural network archi-
Huang et al.18 as a new convolutional network architecture which tecture was Detectron Faster R-CNN with the pre-trained weight
connects each layer to every other layer in a feed-forward fashion. from COCO Detection with Faster R-CNN.
DenseNet incorporates the properties of identity mappings, deep
supervision, reduced feature redundancy, and diverse depth to
enable feature reuse, making it a good feature extractor for vari- 2.5 | Evaluation measures
ous computer vision tasks which improve the accuracy of object
classification. The metrics used to evaluate the machine learning algorithms in
Due to the limited amount of publicly available datasets, we per- bioinformatics were used in this study. 20 To evaluate the perfor-
formed an additional data augmentation to the training data by scal- mance of the DenseNet121 classification method in the classifying
ing, rotation, horizontal flipping, and adjustment of the saturation OSCC apart from the normal oral mucosa on the oral photograph,
19
and exposure. The manually cropped input images of OSCC were we use the precision, recall, F1 score, sensitivity, specificity, and
used to train the network in a minibatch manner. area under the receiver operating characteristics (ROC) curve
Of these 700 photographic images, 490 (70%) were randomly (AUC) to measure the classification performance of the algorithm.
assigned as training data (245 images of OSCC and 245 of normal For the object detection, we evaluated the accuracy of the Faster
oral mucosa), 70 (10%) were used as validation (35 images of OSCC R-CNN to detect a bounding box relative to the ground truth re-
and 35 of normal oral mucosa), and 140 (20%) were used as testing gion in the cancerous images by the precision, recall, F1 score, and
data (70 images of OSCC and 70 of normal oral mucosa) to confirm AUC of precision-recall curve. Furthermore, if the IoU value be-
the accuracy. Fivefold cross-validation was applied to evaluate the tween the generated bounding box and the ground truth was less
robustness of the faster R-CNN performance. than 0.5, then the produced bounding box was considered to be a
For implementation, the experiments are divided into two cate- false prediction as follows:
gories: image classification and object detection. The image classifi- IoU = area of overlap/area of union.
cation experiment was tested on Google Colab using a Tesla P100, Precision = TP/TP + FP = TP/all detections.
Nvidia driver: 460.32, and CUDA: 11.2. The image was prepro- Recall = TP/TP + FN = TP/all ground truths.
cessed by augmentation using Keras ImageDataGenerator; then, the F1 score = 2 x (Precision x Recall)/Precision + Recall.
framework resized an input image to 224 by 224 pixels to feed into Sensitivity = TP/TP + FN.
a deep learning neural network. The neural network architecture Specificity = TN/TN + FP.
914 | WARIN et al.
TA B L E 1 Binary image classification and object detection results “OSCC” vs. “normal oral mucosa”
Models Precision (%) Recall (%) F1 score (%) Sensitivity (%) Specificity (%) AUC of ROC curve
DenseNet121 100.00 99.00 99.00 98.75 100.00 0.99
(ROC curve)
Faster R-CNN 76.67 82.14 79.31 – – 0.79
(precision-recall curve)
Abbreviations: AUC, area under the curve; R-CNN, regional convolutional neural networks; ROC, receiver operating characteristics.
The object detection model trained by the faster R-CNN algorithm
was evaluated on the test set, and the results are shown in Table 1.
The detection of lesions achieved a precision of 76.69%, a recall of
82.14%, and an F1 score of 79.31%. The AUC of the precision-recall
curve also achieved a high accuracy of 0.79 (Figure 4). Figure 5 shows
examples of the output from the object detection model. The fig-
ure shows different samples of the true and false positives detection
of cancerous and non-cancerous from oral images. The analysis was
based on the bounding box generated by the faster R-CNN algorithm
in comparison with the bounding box of the ground truth. One image,
which achieved IoU of 0.49 (less than 0.5), was falsely predicted as a
cancerous lesion by the model (Figure 5H). Figure 5J shows the false
positive detection of a cancerous lesion in a non-cancer image.
F I G U R E 2 The receiver operating characteristic (ROC) curve
of the DenseNet121 classification model showed the AUC of 0.99
4 | DISCUSSION
Oral cancer screening is an examination performed by a general
• True positive (TP): The model classifies as positive, and the actual dental practitioner or physician to look for signs of cancer or pre-
label is also positive which IOU >0.5. cancerous conditions in the oral cavity. In the primary care hospital,
• False positive (FP): The model classifies as positive, but the actual healthcare providers under the consultation of the dentist will help
label is negative which IOU <0.5. with screening and refer patients with suspected oral cancer to the
• True negative (TN): The model classifies as negative, and the ac- cancer center hospital for diagnosis confirmation and treatment. In
tual label is also negative. this work, we proposed the model for automatic classification and
• False negative (FN): The model classifies as negative, but the label detection of cancerous oral lesions in oral photographs that will help
is actually positive. in oral cancer screening. The algorithm used in this study was based
on deep CNN via DenseNet121 to classify oral cancer apart from
the normal oral mucosa and to detect this malignant lesion in pho-
3 | R E S U LT S tographic images via Faster R-CNN. The dataset consisted of pho-
tographic images with biopsy results as the gold standard for oral
The evaluation was performed on the test set, and the results of the cancer diagnosis. All lesion boundaries were identified by three oral
image classification model are reported in Table 1. The identifica- and maxillofacial surgeons as the ground truth for model training,
tion of images which contained cancerous lesions achieved a preci- validation, and testing.
sion of 100%, a recall of 99%, an F1 score of 98.75%, a sensitivity The classification performance of our study is relatively high as
of 99%, and a specificity of 100%. The AUC of the ROC curve of seen in the result which achieved a precision of 100%, a recall of
this model achieved a relatively high accuracy of 0.99 (Figure 2). The 99%, an F1 score of 98.75%, a sensitivity of 99%, a specificity of
model's performance was also evaluated by generating a heat map 100%, and AUC of ROC curve of 0.99. To our knowledge, there are
visualization using the gradient-weighted class activation mapping only 2 studies using a CNN-based classification algorithm to clas-
(Grad-C AM)21 to see how the model identifies the region of interest sify oral cancer from an oral photograph. 22,23 The performance of
and how well the model distinguishes cancer from normal classes. DenseNet121 used in our study was similar to that of the algorithm
Figure 3 shows a sample Grad-C AM output of normal and cancer in the study of Qiuyun et al. 23 but relatively higher than those of
classes and prediction probability. From the output, we observe that ResNet-10 used by the study of Rochan et al. 22 Deep CNNs have
the model perfectly focuses on the key areas of images to classify been used in several studies to classify head and neck lesions and
cancers. have shown acceptable results for classification: for example,
WARIN et al. | 915
F I G U R E 3 The Grad-C AM visualization
of the deep learning-based classification
model. The model correctly classified
a cancerous image input (A) with high
probability of 0.99 and labeled the correct
location (B). The model correctly classified
a non-cancer image input (C) with high
probability of 0.98 and labeled the correct
location (D)
improving information flow and gradients through the network and
could reduce overfitting on tasks with small data training set size.
These properties are significant for deep learning in medicine which
is most often limited by the dataset size.18 Although oral cancer le-
sions in the oral photograph may be obvious for cancer specialists,
they might be difficult to identify by general physicians or healthcare
providers. Thus, this classification model will be beneficial for oral
cancer screening via telemedicine used by general practitioners in
remote regions.
For oral cancer detection, our study showed a good perfor-
mance for the detection of the lesion on the oral photograph with
a precision of 76.67%, a recall of 82.14%, an F1 score of 79.31%,
and AUC of precision-recall curve of 0.79. The detection perfor-
mance was higher than that of the study by Rochan et al. 22 which
also used Faster R-CNN for detecting oral cancer in photographic
F I G U R E 4 The precision-recall curve of the faster R-CNN images. The difference may be due to the imbalance in class distri-
detection model showed the AUC of 0.79 bution between the number of oral cancer images and normal oral
mucosal images used in their study. Faster R-CNN has been widely
used in object detection in medical and dental images, that is, de-
classification of esophageal adenocarcinoma using VGG-16,11 and tection of lymph node metastases from rectal cancer on magnetic
diagnosis of thyroid nodules in ultrasonography images using CNN- resonance imaging (MRI), 25 identification of lung nodules on chest
24
based classification algorithm. DenseNet121 is one of the lat- X-rays, 26 and detection of periodontal defect on panoramic radio-
est CNN-based classification algorithms which has advantages in graphs.13 Currently, there are other CNN models used for object
916 | WARIN et al.
F I G U R E 5 Bounding box ground truth
based on experts annotation and the
output from the faster R-CNN when using
fivefold cross-validation from different
oral photographic images showing correct
prediction in (B), (D), and (F) and a false
prediction in (H) and (J)
WARIN et al. | 917
detection in medical images that show promising results; that is, ORCID
the Mask R-CNN model used for the detection of a lung tumor on Kritsasith Warin [Link]
a positron emission tomography (PET) image, 27 the You Only Look Siriwan Suebnukarn [Link]
Once (YOLO) model used for the detection of a dental restoration in
an oral photograph, 28 and the Single Shot MultiBox Detector (SSD) REFERENCES
model used for the detection of temporomandibular joint arthritis in 1. Sung H, Ferlay J, Siegel RL, et al. Global cancer statistics 2020:
a panoramic radiograph. 29 GLOBOCAN estimates of incidence and mortality worldwide for
36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209-249.
Our study also had one of the common deep learning problems in
2. Warnakulasuriya S. Global epidemiology of oral and oropharyngeal
medicine, which was the limited amount of data.7 This small dataset cancer. Oral Oncol. 2009;45(4-5):309-316.
of oral cancer photographs is due to the fact that the data came from 3. Shah JP, Gil Z. Current concepts in management of oral cancer–
only one cancer center. In the training dataset to develop the model, surgery. Oral Oncol. 2009;45(4-5):394-4 01.
4. Chen SW, Zhang Q, Guo ZM, et al. Trends in clinical features and
we solved this problem by using a data augmentation method.19
survival of oral cavity cancer: fifty years of experience with 3,362
As we know, the key for deep learning is a large dataset which will consecutive cases from a single institution. Cancer Manag Res.
improve the performance model. For future study, advanced aug- 2018;10:4523-4535.
mentation techniques and recent deep CNN algorithms should be 5. Carrard VC, Roxo Gonçalves M, Rodriguez Strey J, et al. Telediagnosis
of oral lesions in primary care: The EstomatoNet Program. Oral Dis.
implemented to further improve the model performance. The data-
2018;24(6):1012-1019. [Link]
set could be expanded by collecting cases from multi-cancer centers 6. Kar A, Wreesmann VB, Shwetha V, et al. Improvement of oral can-
and gathering more image data from an oral cancer screening pro- cer screening quality and reach: the promise of artificial intelli-
gram via telemedicine. gence. J Oral Pathol Med. 2020;49:727-730.
In conclusion, we found that the combination of DenseNet121 7. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning
in healthcare. Nat Med. 2019;25:24-29.
and the Faster R-CNN algorithm gives acceptable results for the
8. Huang ML, Lin TY. Dataset of breast mammography images with
classification and detection of oral cancer in oral photography. masses. Data Brief. 2020;31:105928.
Therefore, we expect that this research will be a fundamental model 9. Zhao Z, Zheng P, Xu S, Wu X. Object detection with deep learning:
for the development of the application for oral cancer screening, a review. IEEE Trans Neural Netw Learn Syst. 2019;30:3212-3232.
10. Ren S, He K, Girshick R, Sun J. Faster R-CNN: towards real-time
especially through telemedicine, which will help inexperienced clini-
object detection with region proposal networks. IEEE Trans Pattern
cians and healthcare providers in remote areas to detect malignancy Anal Mach Intell. 2017;39:1137-1149.
lesions in the oral cavity. 11. Ghatwary N, Zolgharni M, Ye X. Early esophageal adenocarcinoma
detection using deep learning methods. Int J Comput Assist Radiol
Surg. 2019;14:611-621.
AC K N OW L E D G M E N T S S
12. He X, Wang S, Shi S, et al. Computer-Aided Clinical Skin Disease
We gratefully acknowledge the support of the Thammasat University Diagnosis Using CNN and Object Detection Models. 2019 IEEE
Research Grant (TUFT24/2564) and the NVIDIA Corporation for the International Conference on Big Data (Big Data). New York:
Titan Xp GPU used in this research. We thank Sothana Vicharueang IEEE; 2019;4839-4844. [Link] a470
90.2019.9006528
and Waranthorn Chansawang for their assistance with the deep
13. Thanathornwong B, Suebnukarn S. Automatic detection of peri-
learning model training. odontal compromised teeth in digital panoramic radiographs using
faster regional convolutional neural networks. Imaging Sci Dent.
C O N FL I C T O F I N T E R E S T 2020;50:169-174.
14. Kuwada C, Ariji Y, Fukuda M, et al. Deep learning systems for de-
All authors declare that they do not have any conflict of interests.
tecting and classifying the presence of impacted supernumerary
teeth in the maxillary incisor region on panoramic radiographs. Oral
AU T H O R C O N T R I B U T I O N Surg Oral Med Oral Pathol Oral Radiol. 2020;130:464-469.
Kritsasith Warin: Conceptualization; Data curation; Formal analysis; 15. Amin MB, Edge S, Greene FL, et al. AJCC Cancer Staging Manual. Vol.
Funding acquisition; Investigation; Methodology; Project admin- 2017, 8th ed. Switzerland: Springer Nature; 2017:1-1032. https://
[Link][Link]/gp/book/9783319406176
istration; Validation; Writing-original draft; Writing-review & edit-
16. Shah J, Patel S, Singh B, Wong R. Jatin Shah's Head and Neck Surgery
ing. Wasit Limprasert: Formal analysis; Methodology; Software; and Oncology. Vol. 2019, 5th ed. Amsterdam: Elsevier; 2019:1-896.
Validation. Siriwan Suebnukarn: Conceptualization; Methodology; [Link] i [Link]/books/ j atin - s hahs -h ead- a nd-n eck-
Supervision; Validation; Visualization; Writing-review & editing. surger y-and-oncology/shah/978-0 -323-41518-7
17. VisionMarker LW, GitHub. 2019. [Link] 7/visio
Suthin Jinaporntham: Resources. Patcharapon Jantana: Software;
nmarker/. Accessbility verified March 4, 2021.
Validation. 18. Huang G, Liu Z, Maaten LVD, Weinberger KQ. Densely Connected
Convolutional Networks. Paper presented at: 2017 IEEE Conference
E T H I C A L A P P R OVA L on Computer Vision and Pattern Recognition (CVPR). New York:
IEEE; 2017:2261-2269. [Link]
This study was approved by the institute research ethics commit-
19. Shorten C, Khoshgoftaar TM. A survey on Image Data Augmentation
tee (COE No. 020/2563) in accordance with the 1964 Helsinki for Deep Learning. Journal of Big Data. 2019;6(1). [Link]
Declaration and its later amendments. org/10.1186/s40537-019-0197-0
918 | WARIN et al.
20. Jiao Y, Du P. Performance measures in evaluating machine learn- 27. Zhang R, Cheng C, Zhao X, Li X. Multiscale mask R-CNN-
ing based bioinformatics predictors for classifications. Quantitative based lung tumor detection using PET imaging. Mol Imaging.
Biology. 2016;4:320-330. 2019;18:1536012119863531.
21. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. 28. Takahashi T, Nozaki K, Gonda T, Mameno T, Ikebe K. Deep learning-
Grad-C AM: Visual explanations from deep networks via gradient- based detection of dental prostheses and restorations. Sci Rep.
based localization. Int J Comput Vis. 2020;128(2):336-359. http:// 2021;11:1960.
[Link]/10.1007/s11263- 019- 01228-7 29. Lee KS, Kwak HJ, Oh JM, et al. Automated detection of TMJ
22. Welikala RA, Remagnino P, Lim JH, et al. Automated detection and osteoarthritis based on artificial intelligence. J Dent Res.
classification of oral lesions using deep learning for early detection 2020;99:1363-1367.
of oral cancer. IEEE Access. 2020;8:132677-132693. [Link]
org/10.1109/access.2020.3010180
23. Fu Q, Chen Y, Li Z, et al. A deep learning algorithm for detection of
How to cite this article: Warin K, Limprasert W,
oral cavity squamous cell carcinoma from photographic images: a
Suebnukarn S, Jinaporntham S, Jantana P. Automatic
retrospective study. EClinicalMedicine. 2020;27:100558.
24. Koh J, Lee E, Han K, et al. Diagnosis of thyroid nodules on ultra- classification and detection of oral cancer in photographic
sonography by a deep convolutional neural network. Sci Rep. images using deep learning algorithms. J Oral Pathol Med.
2020;10:15245. 2021;50:911–918. [Link]
25. Ding L, Liu G, Zhang X, et al. A deep learning nomogram kit for
predicting metastatic lymph nodes in rectal cancer. Cancer Med.
2020;9:8809-8820.
26. Su Y, Li D, Chen X. Lung nodule detection based on faster R-CNN
framework. Comput Methods Programs Biomed. 2021;200:105866.
This document is a scanned copy of a printed document. No warranty is given about the
accuracy of the copy. Users should refer to the original published version of the material.