Tooth segmentation from panoramic radiography images with deep learning: comparison of UNet, FPN and PSPNet

Salih Taha Alperen Özçelik; Hüseyin Üzen; Abdülkadir Şengür; Adalet Çelebi

Article Details

Tooth segmentation from panoramic radiography images with deep learning: comparison of UNet, FPN and PSPNet

Authors:

Salih Taha Alperen Özçelik | Hüseyin Üzen | Abdülkadir Şengür | Adalet Çelebi

Publisher:

Nigde Omer Halisdemir University

Main Topics:

Görüntü İşleme | Derin Öğrenme | Yapay Zeka (Diğer)

Journal:

Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi Vol.13 Issue 4, pp.1347-1354

Document Type:

Article

Publication Date:

10/2024

Document Language:

Turkish

Document ID (Org No.):

2001329711

pISSN:

2564-6605

Keywords:

Abstract:

 Panoramic radiography is a widely used tool for the early diagnosis and treatment of dental diseases. However, the success of these techniques depends on the dentists experience and ability to interpret correctly. In this study, dental segmentation was performed using deep learning on panoramic radiography images to reduce human error and provide faster solutions. A total of 313 panoramic radiographs from patients aged 18-65 were pixel-wise labeled by an experienced oral and maxillofacial surgeon. Segmentation was performed using UNet, FPN, and PSPNet architectures, and the UNet model with a VGG16 backbone achieved the best result with a 93.74 F1 score, 88.22 IoU, and 98.25 accuracy. Additionally, tests were conducted with EfficientNet, ResNet50, InceptionV3, DenseNet121, and MobileNet backbones, with EfficientNet showing the highest performance with a 93.91 F1 score, 88.52 IoU, and 98.26 accuracy.

 Panoramic radiography is a widely used tool for the early diagnosis and treatment of dental diseases. However, the success of these techniques depends on the dentists experience and ability to interpret correctly. In this study, dental segmentation was performed using deep learning on panoramic radiography images to reduce human error and provide faster solutions. A total of 313 panoramic radiographs from patients aged 18-65 were pixel-wise labeled by an experienced oral and maxillofacial surgeon. Segmentation was performed using UNet, FPN, and PSPNet architectures, and the UNet model with a VGG16 backbone achieved the best result with a 93.74 F1 score, 88.22 IoU, and 98.25 accuracy. Additionally, tests were conducted with EfficientNet, ResNet50, InceptionV3, DenseNet121, and MobileNet backbones, with EfficientNet showing the highest performance with a 93.91 F1 score, 88.52 IoU, and 98.26 accuracy.

 Panoramic radiography is a widely used tool for the early diagnosis and treatment of dental diseases. However, the success of these techniques depends on the dentists experience and ability to interpret correctly. In this study, dental segmentation was performed using deep learning on panoramic radiography images to reduce human error and provide faster solutions. A total of 313 panoramic radiographs from patients aged 18-65 were pixel-wise labeled by an experienced oral and maxillofacial surgeon. Segmentation was performed using UNet, FPN, and PSPNet architectures, and the UNet model with a VGG16 backbone achieved the best result with a 93.74 F1 score, 88.22 IoU, and 98.25 accuracy. Additionally, tests were conducted with EfficientNet, ResNet50, InceptionV3, DenseNet121, and MobileNet backbones, with EfficientNet showing the highest performance with a 93.91 F1 score, 88.52 IoU, and 98.26 accuracy.

 Panoramik radyografi, diş hastalıklarının erken teşhis ve tedavisinde yaygın olarak kullanılan bir araçtır. Ancak bu tekniklerin başarısı, diş hekiminin deneyimine ve doğru yorum yapma yeteneğine bağlıdır. Bu çalışmada, insan hatasını azaltmak ve hızlı çözümler sunmak amacıyla panoramik radyografi görüntülerinden derin öğrenme ile diş segmentasyonu yapılmıştır. 18-65 yaş arasındaki hastalardan alınan 313 panoramik radyografi görüntüsü, deneyimli bir ağız ve çene cerrahı tarafından piksel seviyesinde etiketlenmiştir. UNet, FPN ve PSPNet mimarileri ile segmentasyon yapılmış, VGG16 omurga ağı ile UNet modeli 93,74 F1 skoru, 88.22 KüB ve 98.25 doğruluk ile en iyi sonucu vermiştir. Ayrıca EfficientNet, ResNet50, InceptionV3, DenseNet121 ve MobileNet omurga ağları ile denemeler yapılmış, en yüksek performans EfficientNet omurga ağı ile 93.91 F1 skoru, 88.52 KüB ve 98.26 doğruluk olarak elde edilmiştir.

Article Details

Tooth segmentation from panoramic radiography images with deep learning: comparison of UNet, FPN and PSPNet

Authors:

Salih Taha Alperen Özçelik | Hüseyin Üzen | Abdülkadir Şengür | Adalet Çelebi

Publisher:

Nigde Omer Halisdemir University

Main Topics:

Görüntü İşleme | Derin Öğrenme | Yapay Zeka (Diğer)

Journal:

Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi Vol.13 Issue 4, pp.1347-1354

Document Type:

Article

Publication Date:

10/2024

Document Language:

Turkish

Document ID (Org No.):

2001329711

pISSN:

2564-6605

Deep Learning

Tooth Segmentation

UNetDerin Öğrenme

Diş segmentasyonu

UNet

Issues

Article Details

Tooth segmentation from panoramic radiography images with deep learning: comparison of UNet, FPN and PSPNet

Authors: Salih Taha Alperen Özçelik | Hüseyin Üzen | Abdülkadir Şengür | Adalet Çelebi

Publisher: Nigde Omer Halisdemir University

Main Topics: Görüntü İşleme | Derin Öğrenme | Yapay Zeka (Diğer)

Journal: Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi Vol.13 Issue 4, pp.1347-1354

Document Type: Article

Publication Date: 10/2024

Document Language: Turkish

Document ID (Org No.): 2001329711

pISSN: 2564-6605

Deep Learning

Tooth Segmentation

UNetDerin Öğrenme

Diş segmentasyonu

UNet

Issues

Authors:

Salih Taha Alperen Özçelik | Hüseyin Üzen | Abdülkadir Şengür | Adalet Çelebi

Publisher:

Nigde Omer Halisdemir University

Main Topics:

Görüntü İşleme | Derin Öğrenme | Yapay Zeka (Diğer)

Journal:

Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi Vol.13 Issue 4, pp.1347-1354

Document Type:

Article

Publication Date:

10/2024

Document Language:

Turkish

Document ID (Org No.):

2001329711

pISSN:

2564-6605