| Sign up

PDF (5.4 MB)

Cite

Collect

Submit Manuscript

Show Outline

Figures (6)

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Tables (7)

Table 1

Table 2

Table 3

Table 4

Table 5

Open Access

GFDet: Multi-Level Feature Fusion Network for Caries Detection Using Dental Endoscope Images

Nan Gao^¹(), Yukai Li^¹, Peng Chen^¹, Jijun Tang^², Tianshuang Liu^³()

1College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou 310000, China

2Department of Computer Science and Engineering, University of South Carolina, Columbia, SC 29208, USA

3Future Science and Technology City Branch, Hangzhou Stomatology Hospital, Hangzhou 310000, China

Show Author Information

Abstract

Early dental caries detection by endoscope can prevent complications, such as pulpitis and apical infection. However, automatically identifying dental caries remains challenging due to the uncertainty in size, contrast, low saliency, and high interclass similarity of dental caries. To address these problems, we propose the Global Feature Detector (GFDet) that integrates the proposed Feature Selection Pyramid Network (FSPN) and Adaptive Assignment-Balanced Mechanism (AABM). Specifically, FSPN performs upsampling with the semantic information of adjacent feature layers to mitigate the semantic information loss due to sharp channel reduction and enhance discriminative features by aggregating fine-grained details and high-level semantics. In addition, a new label assignment mechanism is proposed that enables the model to select more high-quality samples as positive samples, which can address the problem of easily ignored small objects. Meanwhile, we have built an endoscopic dataset for caries detection, consisting of 1318 images labeled by five dentists. For experiments on the collected dataset, the F1-score of our model is 75.6%, which out-performances the state-of-the-art models by 7.1%.

Keywords

caries detection feature fusion label assignment object feature enhancement channel attention

References

[1]

C. Valkenburg, F. A. van der Weijden, and D. E. Slot, Plaque control and reduction of gingivitis: The evidence for dentifrices, Periodontology 2000, vol. 79, no. 1, pp. 221–232, 2019.

Crossref Google Scholar

[2]

S. Sälzer, C. Graetz, C. E. Dörfer, D. E. Slot, and F. A. Van der Weijden, Contemporary practices for mechanical oral hygiene to prevent periodontal disease, Periodontology 2000, vol. 84, no. 1, pp. 35–44, 2020.

Crossref Google Scholar

[3]

J. E. Frencken, P. Sharma, L. Stenhouse, D. Green, D. Laverty, and T. Dietrich, Global epidemiology of dental caries and severe periodontitis–A comprehensive review, J. Clin. Periodontol., vol. 44, no. S18, pp. S94–S105, 2017.

Crossref Google Scholar

[4]

M. A. Peres, L. M. D. Macpherson, R. J. Weyant, B. Daly, R. Venturelli, M. R. Mathur, S. Listl, R. K. Celeste, C. C. Guarnizo-Herreño, C. Kearns, et al., Oral diseases: A global public health challenge, Lancet, vol. 394, no. 10194, pp. 249–260, 2019.

Crossref Google Scholar

[5]

Y. Miki, C. Muramatsu, T. Hayashi, X. Zhou, T. Hara, A. Katsumata, and H. Fujita, Classification of teeth in cone-beam CT using deep convolutional neural network, Comput. Biol. Med., vol. 80, pp. 24–29, 2017.

Crossref Google Scholar

[6]

J. Krois, T. Ekert, L. Meinhold, T. Golla, B. Kharbot, A. Wittemeier, C. Dörfer, and F. Schwendicke, Deep learning for the radiographic detection of periodontal bone loss, Sci. Rep., vol. 9, no. 1, p. 8495, 2019.

Crossref Google Scholar

[7]

R. Li, J. Zhu, Y. Wang, S. Zhao, C. Peng, Q. Zhou, R. Sun, A. Hao, S. Li, Y. Wang, et al., Development of a deep learning based prototype artificial intelligence system for the detection of dental caries in children, (in Chinese), Chin. J. Stomatol., vol. 56, no. 12, pp. 1253–1260, 2021.

Google Scholar

[8]

T. Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, Feature pyramid networks for object detection, in Proc. IEEE Conf. Computer Vision and Pattern Recognition, Honolulu, HI, USA, 2017, pp. 936– 944.

Model	Precision	Recall	F1	F2
Faster-RCNN^[41]	54.8	65.0	59.5	62.7
FCOS- $P_{5}$ ^[25]	50.3	70.7	58.8	65.4
GFL-R50^[9]	55.4	70.8	62.2	67.1
Swin-T^[42]	55.7	74.7	63.8	70.0
PVT-tiny^[37]	57.4	84.8	68.5	77.4
Ours	64.1	92.1	75.6	78.4

Dataset	Model	Precision	Recall	F1	F2
SSD-GPNet	SSD	78.4	88	82.9	85.9
	PVT	80.1	94.6	86.7	91.3
	Ours	83.5	98.9	90.5	95.4
CVC-ClinicDB	Mask-RCNN^[44]	83.5	93	88	90.9
	Faster-RCNN^[45]	86.6	98.5	92.2	95.9
	CenterNet-104^[29]	95.3	93.4	94.3	93.7
	PVT^[37]	94.7	96.9	95.8	96.5
	Ours	95.6	97.7	96.6	97.3
CVC-ColonDB	WE-SVM^[50]	50.3	86.3	63.6	75.5
	MDeNetplus^[47]	88.4	91	90	90.5
	COTR^[48]	91.7	93.5	92.6	93.1
	PVT^[37]	94.4	96.1	95.2	95.8
	Ours	95.9	98.7	97.3	98.1

Mechanism	Precision	Recall	F1
Max-IoU	57.4	86.8	69.1
ATSS (topk = 9)^[56]	61.0	90.6	72.9
ATSS (topk = 15)^[56]	60.3	90.7	72.4
AABM	61.8	90.9	73.6

$α$	Precision	Recall	F1
0.1	62.1	90.6	73.7
0.2	63.2	90.9	74.6
0.3	64.1	92.1	75.6
0.4	63.1	91.7	74.8
0.5	62.5	91.2	74.2

Algorithm 1　Label assignment algorithm
Input:
$I :$ Input image;
$B$ : Set of ground truth bounding boxes $b_{i}$ ;
$A$ : Set of anchors $a_{j}$ in image;
$X$ : Number of feature pyramid levels;
$L (θ)$ : Loss function;
$f$ : Hyper-parameter about number of positive anchors
Output:
$P$ : Set of positive samples;
$N$ : Set of negative samples
1: for each ground truth $b_{i} \in B$ do
2:　Build an empty set for candidate positive samples of the 　　ground truth $b_{i} : C_{b} \leftarrow \emptyset$ ;
3:　for each level $\in [1, X]$ do
4:　　Compute loss $L (θ)$ between $a_{j}$ and $b_{i}$ with Eq. (8);
5:　　 $S \leftarrow$ select f anchors from A based on $L (θ)$ ;
6:　　 $C_{b} = C_{b} \cup S$ ;
7:　end for
8:　 $P = C_{b}$ ;
9:　 $N = A - P$ ;
10:　Compute IoU between $P$ and $b_{i} : D_{b} = I o U (P, b_{i})$ ;
11:　Compute mean of $D_{b} : m_{b} = M e a n (D_{b})$ ;
12:　Compute standard deviation of $D_{b} : v_{b} = S t d (D_{b})$ ;
13:　Compute IoU threshold for positive samples: $t_{1} = m_{b} - v_{b}$ ;
14:　Compute IoU threshold for negatibe samples: $t_{2} = m_{b} + v_{b}$ ;
15:　for each positive sample $p_{j} \in P$ do
16:　　if $I o U (p_{j}, b_{i}) < t_{1}$ then
17:　　　Delete $p_{j}$ from P;
18:　　end if
19:　end for
20:　for each negative samples $n_{j} \in N$ do
21:　　if $I o U (n_{j}, b_{i}) > t_{2}$ then
22:　　　Delete $n_{j}$ from N;
23:　　end if
24:　end for
25: end for
26: return $P and N$

Model	Precision	Recall	F1
FPN	57.4	84.8	68.5
FPN + DeConv^[51]	58.6	86.5	69.9
FPN + CARAFE^[38]	59.3	90.4	71.6
FPN + GFU	61.1	90.9	73.1
FPN + SE^[52]	59.0	89.4	71.1
FPN + GFS	60.5	90.5	72.5

Model	Precision	Recall	F1
PAFPN^[53]	59.1	87.9	70.7
NAS-FPN^[54]	58.4	90.4	71.0
DyHead^[55]	59.9	90.9	72.2
FSPN (GFU+GFS)	62.4	91.8	74.3

Mechanism	Precision	Recall	F1
Top1	63.1	91.4	74.7
Top2	63.4	91.7	75.0
Top3	63.5	91.8	75.1
Top4	64.1	92.1	75.6
Top5	63.2	90.8	74.5