A MapReduce-like Deep Learning Model

for the Depth Estimation of Periodontal Pockets

Yusuke Moriyama

, Chonho Lee

, Susumu Date

, Yoichiro Kashiwagi

, Yuki Narukawa

Kazunori Nozaki

and Shinya Murakami

Graduate School of Information Science and Technology, Osaka University, Osaka, Japan

Cybermedia Center, Osaka University, Osaka, Japan

Graduate School of Dentistry, Osaka University, Osaka, Japan

Osaka University Dental Hospital, Osaka, Japan

{yoichiro, y narukawa, ipshinya}@dent.osaka-u.ac.jp, knozaki@dent.osaka-u.ac.jp

Keywords:

Periodontal Disease, Periodontal Pocket, Convolutional Neural Networks, Deep Learning, Object Detection.

Abstract:

This paper explores the feasibility of diagnostic imaging using a deep learning-based model, applicable to

periodontal disease, especially periodontal pocket screening. Having investigated conventional approaches,

we ﬁnd two difﬁculties to estimate the pocket depth of teeth from oral images. One is the feature extraction of

Region of Interest (ROI), which is pocket region, caused by the small ROI, and another is tooth identiﬁcation

caused by the high heterogeneity of teeth (e.g., in size, shape, and color). We propose a MapReduce-like

periodontal pocket depth estimation model that overcomes the difﬁculties. Speciﬁcally, a set of MapTasks

is executed in parallel, each of which only focuses on one of the multiple views (e.g., front, left, right, etc.)

of oral images and runs an object detection model to extract the high-resolution pocket region images. After

a classiﬁer estimates pocket depth from the extracted images, ReduceTasks aggregate the pocket depth with

respect to each pocket. Experimental results show that the proposed model effectively works to achieve the

estimation accuracy to 76.5 percent. Besides, we verify the practical feasibility of the proposed model with

91.7 percent accuracy under the condition that a screening test judges severe periodontitis (6 mm or more).

1 INTRODUCTION

Deep learning algorithms such as convolutional neu-

ral networks (CNN) (LeCun et al., 2015), have

rapidly become a promising methodology for diag-

nostic imaging. This methodology can be used to au-

tomatically diagnose the presence of tuberculosis in

chest radiographs (Ting et al., 2018), to detect mac-

ular degeneration from fundus images (Burlina et al.,

2017), and to locate malignant melanoma in skin im-

ages (Esteva et al., 2017).

Such diagnostic imaging is also important in the

ﬁeld of dentistry due to increasing demands on den-

tal health care. Periodontal disease is one of the most

pervasive infections in the world and reported 80 per-

cent adults are infected with. It is mainly caused of

tooth loss by the chronic periodontal inﬂammation

resulting in the destruction of tooth support tissues.

Moreover, it has recently become apparent that the

disease may be one of the risk factors of diabetes and

stroke (Khader et al., 2006)(Wu et al., 2000). The dis-

ease progresses chronically over a number of years, so

usually, patients have no clear symptom until patho-

logical severe condition resulting in loosening or loss

of teeth. To notice the disease in its early stages that

the patients have rare subjective symptoms is effec-

tive for preventing the progression. An application

that people easily self-check the risk of severity with-

out imparting burdens enhances patient participation

in dental consultations.

This paper explores the feasibility of diagnostic

imaging applicable to periodontal disease. Speciﬁ-

cally, we design a deep learning model to estimate

the depth of periodontal pockets that dentists gener-

ally measures with a probe to examine the severity of

periodontal disease.

In order to efﬁciently extract and study the disease

conditions from the image, we focused on two data

features, hierarchy and regression, observed in oral

images. Hierarchy means that teeth are numbered 1

to 8 in the up and down, left and right, and there are

388

Moriyama, Y., Lee, C., Date, S., Kashiwagi, Y., Narukawa, Y., Nozaki, K. and Murakami, S.

A MapReduce-like Deep Learning Model for the Depth Estimation of Periodontal Pockets.

DOI: 10.5220/0007405703880395

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 388-395

ISBN: 978-989-758-353-7

6 measuring points in each tooth. Regression means

that the 6 measurement points represent the state of

the tooth, 1 to 8 teeth indicate the state of the jaw, and

the four regions (upper left, upper right, downer left

and downer right) represent the oral cavity.

Due to the complex features of the data observed

in oral images, the required accuracy can not be ob-

tained by merely using the image recognition algo-

rithm. In fact, as a preliminary investigation, we have

trained a conventional CNN model on oral images.

However, the estimation accuracy was around 44 per-

cent. There is a need to implement a model to learn

the periodontal disease condition consisting of two

features collectively.

Besides, we looked at two difﬁculties leading to

such low accuracy; a difﬁculty in the Region of In-

terest (ROI) extraction caused by small pocket region

images, and a difﬁculty in tooth identiﬁcation caused

by a high heterogeneity among teeth.

Based on the lessons learned from our prelim-

inary investigation, we propose a MapReduce-like

deep learning model that automatically assesses the

severity of periodontal disease from oral images. It

is a computational framework for data structures hav-

ing the aforementioned features, and it overcomes the

two difﬁculties. Speciﬁcally, the proposed model ex-

tracts pocket region images following tooth identiﬁ-

cation, processed in parallel with respect to each mul-

tiple view of oral images, and then estimates pocket

depth from the extracted pocket region images.

This paper is structured as follows. Section 2 pro-

vides the prior knowledge on periodontal disease for

further discussion later. Section 3 shows the pre-

liminary results when applying a conventional CNN

model, and then clariﬁes the two difﬁculties we tackle

to improve the estimation accuracy. Section 4 de-

scribes the proposed MapReduce-like deep learning

model. Section 5 shows the evaluation results to

verify the effect of the proposed model. Section 6

presents related work and discusses future work.

2 PRIOR KNOWLEDGE ON

PERIODONTAL DISEASE

Periodontal disease or Pyorrhea is a chronic inﬂam-

mation affecting the gums surrounding the teeth. In

its early stage called Gingivitis, the gums become

swollen and may bleed. In an advanced stage of gin-

givitis called Periodontitis, the teeth can become de-

tached from the gums and loosen.

Generally and traditionally, dentists examine the

severity of periodontal disease by measuring the

depth of periodontal pockets for all teeth using a

Lingual

Buccal

Figure 1: Illustration of periodontal pocket measurement

(left) and 6 point measuring point (right).

probe (Figure 1-left). The measured depth is recorded

in millimeters as a pocket chart. Under a common 6-

point method, dentists measure six pockets: the distal

buccal (P1), central buccal (P2), medial buccal (P3),

medial lingual (P4), central lingual (P5), and distal

lingual (P6) (Yoshie et al., 2013), as illustrated in Fig-

ure 1-right. In this paper, P1∼P6 are called pocket

numbers and the gum region surrounding the pockets

is called the pocket region.

In this paper, we deﬁne tooth number with three

letters, for example, as in UL1 and DR8. The ﬁrst

letter is U or D for upper and downer, the second letter

is L or R for left and right from the patient’s view, and

the third letter indicates the number from the median,

respectively. For example, if the tooth number is UL1,

UL1 represents the ﬁrst tooth from the median in the

upper left of the patient.

3 PRELIMINARY

INVESTIGATION

To investigate the feasibility of applying deep learn-

ing techniques into the pocket depth estimation, we

trained a conventional CNN model, called VGG-

16 (Simonyan and Zisserman, 2014) with a set of oral

images from 1333 patients acquired at the Osaka Uni-

versity Dental Hospital. The trained model estimates

the depth of 12 pockets, which is the buccal side of 4

upper front teeth. Using 80 percent of the oral images

as training, and 20 percent as the test, we observed

the estimation accuracy of approximately 44 percent

according to precision in millimeters.

In order to improve the accuracy, we have found

ways to overcome the two difﬁculties in Region of

Interest (ROI) feature extraction and tooth identiﬁca-

tion, which led to such low accuracy.

Difﬁculty in ROI Feature Extraction: The ﬁrst

problem is a small ROI in the oral image. According

to dentists, the correlation between different pockets

is weak even for the same teeth. So, for estimating

a pocket depth, we better focus on the corresponding

A MapReduce-like Deep Learning Model for the Depth Estimation of Periodontal Pockets

389

pocket region. However, each pocket region is small

in an oral image, and the “noise” unrelated to the esti-

mation is large. Examples of the noise include teeth,

lips, and other pocket regions. In addition, due to the

high computational cost and large memory size, the

input image size of the CNN model is limited in gen-

eral. The size of the original oral image used in this

study is large (e.g. 1600 × 1200 and 2080 × 1560),

so it is necessary to reduce the image size to adjust

to the size of model input. Therefore, as illustrated in

Figure 2, if the entire oral image is the model input,

the image resolution of ROI becomes lower than that

of ROI in the original image.

Difﬁculty in Tooth Identiﬁcation: The second

problem is tooth heterogeneity. We focus on a pocket

region to estimate a pocket depth, so we need to iden-

tify the corresponding teeth to extract the pocket re-

gion images. The training dataset contains a set of

oral images taken from different directions for each

of the patients. Depending on the direction, tooth fea-

tures such as color, size and shape vary in those im-

ages even for the same teeth. In the case of general

objects such as dog, cat, etc., that have large feature

differences, the direction is not a problem. However,

in the case of teeth, since the features are inherently

similar, it becomes difﬁcult to identify the teeth.

4 DESCRIPTION OF THE

PROPOSED MODEL

This section describes the proposed MapReduce-like

periodontal pocket depth estimation model together

with the design principles. Figure 3 illustrates the

overview ﬂow of the proposed model, where CNN

pre-processing and post-processing are added. To

overcome the ﬁrst difﬁculty in ROI feature extraction,

we extract high resolution pocket region images, pro-

vided to CNN. To overcome the second difﬁculty in

tooth identiﬁcation, we parallelize the pocket region

extraction. The details of the process are explained in

the following subsections.

4.1 Design Principles

First, we consider the extraction of pocket region im-

ages. To extract the pocket region images, we fur-

ther need to identify tooth numbers which represent

tooth location and pocket numbers, which represent

the pocket location of the teeth. In order to identify

the tooth numbers, we perform tooth detection ﬁrst.

A few methods exist for the detection of teeth,

such as image processing, segmentation, and object

Figure 2: Quality of pocket region image (i.e., ROI). Note

that we consider the CNN accepts a size of 128 ×128 input.

The resolution of ROI input to the CNN is lower than that

of ROI in the original image, and the information is lost.

Figure 3: Overview ﬂow of the pocket depth estimation.

detection. In image processing, teeth can be detected,

but it is difﬁcult to classify tooth numbers. This is

because the tooth numbers must be identiﬁed accord-

ing to the coordinates of the detected teeth. How-

ever, image processing does not work when the align-

ment of the teeth is bad or the teeth are missing. The

segmentation approach also is not suitable because

when specifying tooth coordinates, this approach re-

quires a contour acquisition process after pixel classi-

ﬁcation, which increases labor and processing time.

For these reasons, we use object detection, named

YOLOv2 (Redmon and Farhadi, 2017), which makes

it possible to identify tooth numbers and acquire the

coordinate of teeth. By obtaining teeth coordinates,

we can extract sub-images of the pocket regions from

the oral images based on the teeth coordinates.

Secondly, we parallelize the pocket region extrac-

tion process on the per-viewpoint of oral images. In

other words, each pocket region extraction process

uses oral images with one of the multiple viewpoints.

This way of processing eliminates the necessity of

learning tooth features in multiple directions with one

HEALTHINF 2019 - 12th International Conference on Health Informatics

390

Figure 4: Pocket region extraction process (MapTask).

YOLOv2, and limits the range of tooth numbers. For

example, in the oral image taken from the left direc-

tion, the right teeth are not included, so the classiﬁca-

tion can be limited to only the left teeth. As a result,

we can reduce the misclassiﬁcation of tooth numbers,

which leads to an improvement in accuracy.

4.2 Design of the Proposed Model

In this subsection, we describe the proposed

MapReduce-like periodontal pocket depth estimation

model. The proposed model is divided mainly into

three parts: the Mapping phase, the CNN phase,

and the Reducing phase. The Mapping phase identi-

ﬁes tooth numbers and extracts pocket region images

from oral images. The CNN phase estimates pocket

depth from the pocket region images. The Reducing

phase aggregates the estimated depth with respect to

each identical pocket.

Mapping Phase: In this phase, a set of MapTasks

that recognize teeth and extract pocket region images

from oral images, is executed in parallel. Figure 4

shows our proposed pocket region extraction process

that addresses the difﬁculty in ROI feature extrac-

tion. The process basically uses YOLOv2, known as

a model capable of high-precision, high-speed object

detection, to obtain the high resolution pocket region

images.

YOLOv2 learns the features of each tooth so that

it can detect the teeth and identify the tooth number,

as illustrated in the red dotted box of Figure 4. Since

three pocket regions exist at even intervals as shown

in Figure 1-right, each of the detected tooth images

is equally divided into three so as to correspond to

those three pockets. Then, discarding the lower part

of the tooth and including the gum, we can obtain the

rectangular sub-images of pocket regions, which are

a high resolution of large ROI images.

However, it is difﬁcult for YOLOv2 to identify

teeth with the same number (i.e., similar teeth such as

upper front teeth, UL1 and UR1) because YOLOv2

recognizes objects by looking at features of size,

color, and shape, but not its location. Therefore, in

the case of the front view of the image, we compare

the x-coordinate of the center of the detected rectan-

gular to classify teeth into the left and right.

Each MapTask deals with one of the multiple

viewpoints of oral images. Before parallelization, all

oral images are input to the same MapTask. In the

proposed method, the oral images are sorted in ad-

vance for each shooting direction and entered into

the corresponding MapTask. Each MapTask extracts

pocket region and passes them to the CNN phase.

CNN Phase: In this phase, the CNN classiﬁer esti-

mates pocket depth from the pocket region images.

We use VGG-16 (Simonyan and Zisserman, 2014)

pretrained by ImageNet. The CNN classiﬁer takes

an input image with the size of 128 × 128 and out-

puts one pocket depth with probability as the estima-

tion conﬁdence. Speciﬁcally, pocket region images

extracted in the previous phase are sorted and grouped

by tooth number and pocket number, as identical

pocket region images. A group of identical pocket

region images is input one by one to CNN classiﬁer

and outputs the group of estimated pocket depth.

Reducing Phase: In this phase, a set of ReduceTasks

aggregates the estimated depth. Each ReduceTask

handles one of the groups of pocket depth. There are

a few ways to aggregate the pocket depth estimates

with conﬁdence. For example, the ﬁnal pocket value

is given as one pocket depth estimate with the highest

conﬁdence, or the weighted sum of the pocket depth

estimate and conﬁdence. In this paper, the former

method is adopted. Figure 5 shows the detailed de-

sign of the proposed model.

5 EVALUATION

This section ﬁrst explains the experimental environ-

ment and veriﬁes the effects of the proposed model in

terms of the estimation accuracy of pocket depth by

comparing with that of a conventional CNN model.

Dataset: We use a dataset of 2625 oral images ob-

tained from 1333 patients visiting at Osaka Univer-

A MapReduce-like Deep Learning Model for the Depth Estimation of Periodontal Pockets

391

Figure 5: Design of the proposed model.

sity Dental Hospital. The images include ones taken

from front and left directions. Data labels are the

pocket depth of the teeth, which are measured by a

few trained physicians. In this work, we focus on 12

pockets such as P1, P2, and P3 of teeth UL1, UL2,

UR1 and UR2. The 80 percent of the data is used for

training, and the remaining 20 percent is used for test.

Experimental Setting: The proposed model uses a

CNN model, VGG-16, in its Reducing phase. The

input image size is set to 128 × 128. To conﬁrm fea-

sibility according to usage, we change the number of

classes of CNN based on pocket depth. We conduct

three experiments as follows.

• E1: Screening by 2 stages, e.g., “Healthy” (3 mm

or less) and “Unhealthy” (4 mm or more).

• E2: Severity measurement by 3 stages, e.g.,

“Healthy” (3 mm or less), “Moderate periodon-

titis” (4 or 5 mm), and “ Severe periodontitis”

(6 mm or more).

• E3: Depth estimation by 15 stages, i.e., pocket

depth in millimeters between 1 mm and 15 mm.

In each experiment, we compare the proposed

model with a single CNN model that outputs the esti-

mates of 12 pocket depth at once. In order to perform

multi-label classiﬁcation, we parallelize ﬁnal fully-

connected layer of CNN. For example, in the case

of Screening, the CNN has a 24 neurons (2 stages ×

12 pockets) at ﬁnal layer. For the evaluation metrics,

we compute the classiﬁcation accuracy, and mean-

squared error (MSE) of the distance between the true

pocket depth and the estimates.

5.1 Experimental Result

Table 2 shows the comparison result. It shows that

the proposed model outperforms a single CNN model

in the classiﬁcation accuracy with an improvement of

1.2 percent at E1, 2.6 percent at E2, and 3.0 percent at

E3, respectively. The accuracy improvement slightly

increases as the number of classes increases. This in-

dicates that the effect of the proposed MapReduce-

like parallel processing becomes bigger against more

complicated tasks.

In order to verify the effect, we conduct two

additional experiments. The ﬁrst experiment is to

check how the image resolution of ROI (i.e., the

pocket region image) affects accuracy, as shown in

Figure 2. The second experiment is to check how

the MapReduce-like task parallelization contributes

to the improvement of accuracy.

HEALTHINF 2019 - 12th International Conference on Health Informatics

392

Table 1: Pocket depth distribution of dataset consisting of 1333 patients, 2625 oral images and 15312 pocket regions. Note

that the toothless part and the pocket without measurement data are deleted from the dataset.

Depth (mm) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

The number of data 343 4702 6010 1442 504 654 218 101 104 35 34 4 1 1 0

Table 2: Accuracy (%) of pocket depth estimation using a

single CNN and the proposed model (MSE (mm)).

Experiment CNN Proposed model

Screening 75.3 (0.25) 76.5 (0.21)

Severity measurement 70.5 (0.61) 73.1 (0.52)

Depth estimation 44.0 (2.20) 47.0 (2.14)

Figure 6: Accuracy over different image resolution of ROI.

5.1.1 Effect of Pocket Region Extraction

Different from a single CNN that learns oral images,

the proposed model performs pocket region extraction

and learns the pocket region images with less noise.

In this experiment, we change the resolution of im-

ages to be input to CNN. Speciﬁcally, pocket region

images of 64×64, 10 ×10, and 5×5 are used, respec-

tively. This indicates that the images are magniﬁed to

128 × 128, the input size of CNN.

Figure 6 shows the accuracy over different quali-

ties of the pocket region images in the case of E1 and

E2. As expected, the accuracy improves as the im-

age resolution increases. This implies that the pocket

region extraction contributes to improving the clas-

siﬁcation accuracy by focusing on a pocket region,

which has a ROI with a higher resolution compared

to the ROI in the oral image.

5.1.2 Effect of MapTask Parallelization

Different from a single CNN that learns both front and

left views of oral images at once, the proposed model

parallelizes tooth identiﬁcation tasks, each of which

deals only one-view image. In this experiment, the

average precision (AP) of YOLOv2 is compared over

Table 3: Per-class average precision(%) and mean average

precision(%) of YOLOv2. Note that UL1 and UR1 are rep-

resented as U1, and UL2 and UR2 are represented as U2.

View AP of U1 AP of U2 mAP

Front 92.8 88.9 90.9

Left 79.0 77.6 78.3

Front + Left 62.6 62.3 62.5

Figure 7: Accuracy over different pocket depth at E1.

different sets of oral images that contain a front view,

left view, or both front and left views.

Table 3 shows the per-class AP and its mean

(mAP). When the model uses both the front and left

view images at once, the mAP decreases by more

than 15 percent, compared with the case inputting one

view image independently. Therefore, the MapTask

parallelization is very effective.

Finally, we would like to discuss the low accuracy

at Depth estimation (E3). The dramatic drop in accu-

racy could be caused by the unbiased dataset. The

dataset includes a different number of oral images

with different pocket depths, as shown in Table 1.

It seems that the model was not trained well for the

classes with large pocket depth because of the insufﬁ-

cient number of training data. To further improve the

accuracy, we deﬁnitely need more data, i.e., oral im-

ages especially from patients with severe periodontal

disease. Also, it would be beneﬁcial to use other in-

dices in addition to pocket depth such as the presence

or absence of bleeding and blood sugar level to learn

the relationship between oral images and the severity

of periodontal disease.

5.2 Towards Practical Use

Although we verify the effect of pocket region ex-

traction and parallelization, we need to further im-

prove the estimation accuracy towards the practical

periodontal screening. In this subsection, we show

the result of an additional experiment to investigate

the practical feasibility of periodontal screening with

the proposed model.

Figure 7 shows the accuracy over different pocket

depth at E1. As you can see, the accuracy of 4 and

5 mm is relatively lower than the others. We found

that the pockets with 4 or 5 mm depth are wrongly

classiﬁed as the pocket with 3 mm depth. In fact, it

is also hard for dentists to judge the pockets with 4

or 5 mm depth (i.e., moderate periodontitis) from the

A MapReduce-like Deep Learning Model for the Depth Estimation of Periodontal Pockets

393

images, which do not have much visual difference.

We doubt that pocket region images with ambiguous

labels might be included in training dataset.

Thus, we consider ignoring the data (i.e., pocket

region images) of 3, 4 and 5 mm pocket depth and see

if the proposed model can be used for the screening of

severe periodontitis, which differentiates “Healthy”

(2 mm) from “Severe periodontitis” (6 mm or more).

We named the experiment as E1

. If the pocket depth

is more than 6 mm, the use of a surgical approach is

likely to be necessary(Greenstein, 1997). Therefore,

ﬁnding severe periodontitis is very important. Addi-

tionally, we ask a dentist to select data that he can

distinguish whether it is severe in order to completely

eliminate label ambiguity. That is, we perform E1

two different dataset as follows.

• E1

-a: A dataset of the pocket region images ex-

tracted by the Mapping phase of the proposed

model (Healthy: 10449 images, Severe periodon-

titis: 2537 images).

• E1

-b: A dataset of the pocket region images the

dentist selected from E1

-a dataset (Healthy: 1479

images, Severe periodontitis: 809 images).

Table 4 is the confusion matrix of E1

-a and E1

b. From the results of E1

-a, the accuracy is 87.4 per-

cent, true positive rate (TPR) is 93.3 percent, and false

positive rate (FPR) is 29.7 percent. The accuracy im-

proves compared to E1 case (76.5 percent), but FPR

is so high that the model cannot be used for screen-

ing. From the results of E1

-b, the accuracy is 91.7

percent, TPR is 93.2 percent, and FPR is 6.8 percent.

Compared to E1

-a model, the E1

-b model reduces

FPR by approximately 20 percent.

Figure 8 compares ROC curves of E1 and E1

From Figure 8, the area under the ROC curve (AUC)

at E1

-a is 0.917 and the AUC at E1

-b is 0.962, which

is approximately 0.2 larger than the AUC at E1. This

implies that the E1

makes less miss the case with high

periodontal severity.

Through the experiments, we realize that there is

not a strong relationship between pocket depth and the

visual appearance of the pocket; especially the pock-

ets with 4 mm or 5 mm depth (i.e., moderate periodon-

titis). However, the proposed model can judge severe

periodontitis from pocket region images at over 91.7

percent accuracy with 6.8 percent FPR. As the results,

we show the feasibility of screening for ﬁnding pa-

tients with severe periodontal disease. This will en-

courage people to do self-check at home and to see a

doctor at right timing.

Table 4: Confusion matrix (E1

-a / E1

-b).

Actual

Healthy

Actual

Severe periodontitis

Predicted

Healthy

1806 / 206 201 / 18

Predicted

Severe periodontitis

128 / 20 476 / 144

Figure 8: Comparison of ROC curves. Note that the thresh-

old of the ROC curve uses the score for the Healthy class

which is the output of CNN.

6 RELATED WORK AND

FUTURE WORK

Related Work: Deep learning has been applied in

various ﬁelds. One of the advantages in using deep

learning is the ability to automatically extract effec-

tive and domain-speciﬁc features for tasks.

Deep learning has been applied in medical

ﬁelds (Litjens et al., 2017) as well. Various mod-

els based on convolutional neural networks (CNN)

have been proposed, which detect diabetic retinopa-

thy from retinal fundus photographs (Gulshan et al.,

2016), breast cancer in an mammography (Becker

et al., 2017), and brain lesion segmentation in

MRIs (Kamnitsas et al., 2017). Different from that

type of research, our model parallelizes classiﬁcation

tasks, each of which only focuses on the correspond-

ing ROI with less noise.

In the dental ﬁeld, there are many studies related

to the classiﬁcation of dental diseases from X-ray

images (Prajapati et al., 2017) and of dental plaque

from quantitative light-induced ﬂuorescence (QLF)

images (Imangaliyev et al., 2016); however, few stud-

ies focuses on oral images. To the best of our knowl-

edge, this study is the ﬁrst study to investigate the re-

lationship between oral images and pocket depth us-

ing deep learning.

Future Work: To further improve the estimation

accuracy, the following work will be conducted.

First, we need to investigate the pocket regions

around all teeth in addition to just the upper front teeth

(e.g., UL1 and UR1). Different teeth and their pocket

regions might have unique features that are different

from those of the upper front teeth. As shown in Sec-

tion 5.2.2, when we consider various teeth and more

HEALTHINF 2019 - 12th International Conference on Health Informatics

394

images with multiple views, the effect of the proposed

MapReduce like processing becomes much bigger.

Secondly, the proposed model mainly utilizes two

trained models, YOLOv2 for tooth detection and

CNN for pocket depth estimation. These models are

independently trained on different set of images. We

will try to design an end-to-end model by changing

the current model’s layer composition. This end-to-

end model should be able to simultaneously train the

model instantaneously.

Thirdly, in addition to the oral images, additional

information such as X-ray images and blood test re-

sults should contribute to improving the estimation

accuracy. We will work on designing a model that

can handle a multimodal dataset.

7 CONCLUSION

In this paper, we proposed a MapReduce-like pocket

depth estimation model which performed paral-

lel pocket region extraction processing and multi-

directional information aggregation in a Mapping

phase and Reducing phase, respectively. Through the

experiments, we realize that there is not a strong rela-

tionship between pocket depth and the visual appear-

ance of the pocket. So, it is difﬁcult to judge moder-

ate periodontitis with only oral images. However, we

show the feasibility of screening for ﬁnding patients

with severe periodontal disease. The proposed model

can be used for self-check at home as a tool with the

same quality of vision as the dentists.

ACKNOWLEDGMENT

We would like to thank Osaka University Dental Hos-

pital, for setting up the environment for our research

and the medical dataset for the experiments. This

work was supported by Social Smart Dental Hospi-

tal, a collaborative project between Osaka University

and NEC Corp.

REFERENCES

Becker, A. S., Marcon, M., Ghafoor, S., Wurnig, M. C.,

Frauenfelder, T., and Boss, A. (2017). Deep learning

in mammography: diagnostic accuracy of a multipur-

pose image analysis software in the detection of breast

cancer. Investigative radiology, 52(7):434–440.

Burlina, P. M., Joshi, N., Pekala, M., Pacheco, K. D., Fre-

und, D. E., and Bressler, N. M. (2017). Automated

grading of age-related macular degeneration from

color fundus images using deep convolutional neural

networks. JAMA ophthalmology, 135(11):1170–1176.

Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M.,

Blau, H. M., and Thrun, S. (2017). Dermatologist-

level classiﬁcation of skin cancer with deep neural net-

works. Nature, 542(7639):115.

Greenstein, G. (1997). Contemporary interpretation of

probing depth assessments: Diagnostic and therapeu-

tic implications. a literature review. Journal of Peri-

odontology, 68(12):1194–1205.

Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu,

D., Narayanaswamy, A., Venugopalan, S., Widner, K.,

Madams, T., Cuadros, J., et al. (2016). Development

and validation of a deep learning algorithm for de-

tection of diabetic retinopathy in retinal fundus pho-

tographs. JAMA, 316(22):2402–2410.

Imangaliyev, S., van der Veen, M. H., Volgenant, C. M.,

Keijser, B. J., Crielaard, W., and Levin, E. (2016).

Deep learning for classiﬁcation of dental plaque im-

ages. In International Workshop on Machine Learn-

ing, Optimization and Big Data, pages 407–410.

Springer.

Kamnitsas, K., Ledig, C., Newcombe, V. F., Simpson,

J. P., Kane, A. D., Menon, D. K., Rueckert, D., and

Glocker, B. (2017). Efﬁcient multi-scale 3d cnn with

fully connected crf for accurate brain lesion segmen-

tation. Medical image analysis, 36:61–78.

Khader, Y. S., Dauod, A. S., El-Qaderi, S. S., Alkafajei, A.,

and Batayha, W. Q. (2006). Periodontal status of di-

abetics compared with nondiabetics: a meta-analysis.

Journal of diabetes and its complications, 20(1):59–

68.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. nature, 521(7553):436.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., van der Laak, J. A., van

Ginneken, B., and S

anchez, C. I. (2017). A survey

on deep learning in medical image analysis. Medical

image analysis, 42:60–88.

Prajapati, S., Nagaraj, R., and Mitra, S. (2017). Classiﬁca-

tion of dental diseases u sing cnn and transfer learning.

In Proceedings of the 5th International Symposium on

Computational and Business Intelligence (ISCBI).

Redmon, J. and Farhadi, A. (2017). YOLO9000: Better,

faster, stronger. In Proceedings of the 30th IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 6517–6525.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Ting, D. S., Yi, P. H., and Hui, F. (2018). Clinical applica-

bility of deep learning system in detecting tuberculo-

sis with chest radiography. Radiology, 286(2):729.

Wu, T., Trevisan, M., Genco, R. J., Dorn, J. P., Falkner,

K. L., and Sempos, C. T. (2000). Periodontal dis-

ease and risk of cerebrovascular disease: the ﬁrst na-

tional health and nutrition examination survey and

its follow-up study. Archives of Internal Medicine,

160(18):2749–2755.

Yoshie, H., Itou, K., Murakami, S., and Shin, K. (2013).

Clinical periodontal disease. Ishiyaku Publishing, 2

edition.

A MapReduce-like Deep Learning Model for the Depth Estimation of Periodontal Pockets

395