Malmö University Publications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Deep learning for automated alveolar cleft segmentation and bone graft volume estimation in cone-beam computed tomography imaging – a multicenter study
Malmö University, Faculty of Odontology (OD).ORCID iD: 0009-0007-0252-4165
Department of Applied Oral Sciences & Community Dental Care, Faculty of Dentistry, The University of Hong Kong, Prince Philip Dental Hospital, 34 Hospital Road, Sai Ying Pun, Hong Kong.
DeepCare, 10250 Constellation Blvd, Los Angeles, CA 90067, USA.
Department of Oral and Maxillofacial Surgery, Faculty of Dentistry, The University of Hong Kong, Prince Philip Dental Hospital, 34 Hospital Road, Sai Ying Pun, Hong Kong.
Show others and affiliations
2026 (English)In: Oral surgery, oral medicine, oral pathology and oral radiology, ISSN 2212-4403, E-ISSN 2212-4411, Vol. 141, no 3, p. 400-408Article in journal (Refereed) Published
Abstract [en]

Objective: To train and validate a deep learning-based diagnostic tool capable of accurately segmenting the alveolar cleft region and automatically estimating the required bone graft volume using cone-beam computed tomography (CBCT) imaging.

Study Design: Eighty-eight CBCT scans from patients with non-syndromic unilateral clefts were divided into training (n = 45), validation (n = 10), and test (n = 33) sets. Two annotators performed manual segmentations, and the intersection of these served as the ground truth for training 3D U-Net models. The Dice Similarity Coefficient (DSC) was calculated to validate the tool by comparing manual and automated segmentations. Three observers evaluated the resulting deep learning model using 33 CBCT scans and performing subjective assessments in terms of shape and size.

Results: The DSC between the two annotators was 0.66, and between the automated and manual segmentations, 0.78. The observers considered the automated segmentations acceptable in 82%–94% of the cases. The deep learning-based tool took approximately 7 seconds to perform an automated segmentation, while manual segmentation by the annotators required 14 and 6.5 minutes.

Conclusion: The deep learning-based tool that was developed in the present study can accurately perform automated segmentations of unilateral alveolar clefts and estimate the required bone graft volume.

Place, publisher, year, edition, pages
Elsevier, 2026. Vol. 141, no 3, p. 400-408
National Category
Odontology
Identifiers
URN: urn:nbn:se:mau:diva-81081DOI: 10.1016/j.oooo.2025.10.020PubMedID: 41390267Scopus ID: 2-s2.0-105025147915OAI: oai:DiVA.org:mau-81081DiVA, id: diva2:2020111
Funder
Region Skåne, OFRS979540Available from: 2025-12-09 Created: 2025-12-09 Last updated: 2026-02-04Bibliographically approved
In thesis
1. Radiographic examinations in dental care for paediatric patients with special healthcare needs
Open this publication in new window or tab >>Radiographic examinations in dental care for paediatric patients with special healthcare needs
2025 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Children are more sensitive to ionising radiation than adults due to the higher susceptibility of developing organs and tissues, including sensitive organs such as the thyroid gland, which is more easily exposed to radiation on account of children’s smaller head size. The greater number of remaining life years of young persons means that the risk of cumulative exposure is also greater. Children and youth with special healthcare needs (CYSHCN) often require more frequent examinations, which can result in higher exposure to ionising radiation. This thesis focuses on two groups of CYSHCN, those born preterm and those with orofacial clefts.

Preterm-born children account for about 5% of all children born in Sweden each year. The consequences of preterm birth can lead to higher exposure to ionising radiation due to healthcare-related examinations. However, knowledge on whether dental care also contributes to higher exposure, which Paper I analysed, is sparse. A retrospective study, Paper I retrieved the dental records of 311 preterm- and full term-born children and adolescents. No findings of significant differences in number of dental care-related radiographic examinations were found between the two groups. Additionally, although dental behaviour management problems during examinations and/or treatments in childhood were higher in the preterm group, this did not affect their dental status. Prevalence of caries and rate of referrals to a paediatric dental specialist were comparable between the two groups.

Orofacial clefts are among the most common congenital anomalies, with a prevalence of 2 in 1000 live births. Alveolar bone grafting is a surgical procedure, typically performed when the child is between 9 and 12 years of age, in which autogenous bone grafts are transplanted into the cleft area. Radiographic examinations are required before and after alveolar bone grafting, and cone-beam computed tomography (CBCT) is one of the imaging methods that can be used.

Segmentation of the cleft area can be done using CBCT imaging, offering a clear visualisation of the cleft and enabling estimation of the amount of bone graft required preoperatively. Paper II demonstrated this through semi-automated segmentations of 53 CBCT scans of uni- and bilateral clefts. Moreover, CBCT provided important information about the impact of the cleft on surrounding anatomical structures.

Both manual and semi-automated segmentation require the operator to delineate the cleft area in each CBCT slice, a process that is time-consuming and demands advanced software skills. Artificial intelligence is increasingly being incorporated into healthcare and may be useful in addressing these challenges. Thus, Paper III developed a deep learning tool and tested automated segmentation of unilateral clefts. In all, 88 CBCT scans were used for this purpose. The deep learning tool was found to be capable of accurately segmenting the cleft area as well as estimating preoperative bone graft volume.

Low-dose CBCT protocols are an effective way of reducing radiation exposure. Paper IV tested such a protocol in postoperative radiographic assessment of 14 cleft cases. The low-dose protocol was found to provide adequate image quality; however, observers reported lower confidence in their evaluations compared to the standard-dose protocol.

Abstract [sv]

Barn är känsligare för joniserande strålning än vuxna på grund av, bland annat, deras tillväxt av organ och vävnader och att de förväntas ha en längre tid kvar i livet. Deras mindre huvudstorlek gör också att till exempel sköldkörteln lättare hamnar inom exponerade områden inom ramen för odontologisk radiologi. Två olika barngrupper har inkluderats i denna avhandling: för tidigt födda barn, och barn födda med läpp-käk-gom-spalt (LKG).

Varje år föds drygt 5% av alla barn för tidigt i Sverige. Några av konsekvenserna för dessa barn brukar leda till en ökad mängd undersökningar inom sjukhusvården och därmed också en ökad mängd röntgenundersökningar. Det saknas dock kunskap om vilken roll tandvården bidrar till att öka mängden undersökningar med joniserande strålning, vilket analyserades i avhandlingens första delarbete. Journaler, 311 stycken, från tandvården i Skåne, Sverige, granskades. Resultatet visade att de för tidigt födda- inte exponerades för mer joniserade strålning än barn födda efter fullgången graviditet. Dock uppvisade många av barnen problem i samband med undersökningar och/eller behandlingar under uppväxten jämfört med fullgångna barn. De för tidigt födda hade i stort inte ett sämre oralt status, förekomsten av karies och antal remisser till specialister inom pedodonti var likvärdig jämfört med barn födda efter normal tid.

LKG är bland de vanligaste medfödda ansiktsmissbildningarna, med en prevalens på 2 per 1000 födslar i Sverige. Patienterna genomgår många olika behandlingar i tidig ålder. När barnen är mellan 9–12 år, utför man ett kirurgiskt ingrepp, en s.k. dentoalveolär bentransplantation, där spalten fylls ut och stängs. Röntgenundersökningar krävs både före och efter ingreppet, och cone-beam computed tomography (CBCT) är en av de metoderna som kan användas. CBCT- bilder kan användas för att tydliggöra spaltområde och ta fram tredimensionella (3D) modeller. Resultatet kan medverka till att en uppskattning av mängden ben, som behövs till operationen. I andra delarbetet, togs sådana 3D-modeller fram med hjälp av 53 CBCT-undersökningar av spalter som var antingen på en sida om mittlinjen, ensidiga eller på båda sidorna, dubbelsidiga spalter. Det konstateras också att CBCT kan ge ytterligare information om hur spalten påverkar de omgivande anatomiska strukturerna. Den manuella hanteringen av denna process är tidskrävande för tandläkaren som ska utföra det.

Artificiell intelligens (AI) har alldeles nyligen blivit mycket uppmärksammat. Det integrerar alltmer i olika delar av samhället och då även inom hälso- och sjukvård. Även inom tandvård är AI ett område där det bedrivs mycket forskning i nuläget. Ett djupinlärningsbaserat verktyg, AI, utvecklades och testades i det tredje delarbetet och totalt användes 88 CBCT-bilder. Resultaten visade att AI verktyget kunde avgränsa spaltområdet med hög noggrannhet samt uppskatta mängden ben som skulle behövas till en kommande bentransplantationen. Tidsbesparingen var tydlig.

Att minska stråldoser som ges till patienterna är en del av optimeringsprocessen som inom Europeiska unionen är ett lagstadgat krav för att försöka använda en så liten mängd joniserande strålning vid varje undersökningstillfälle. Att utarbeta och testa s.k. låg-dos CBCT protokoll är ett effektivt sätt att spara, minimera, strålningsdos till patienter. I det fjärde delarbetet testades ett sådant protokoll i den postoperativa röntgenundersökningen av 14 spaltfall. Det visade sig att protokollet gav tillräcklig bildkvalitet för en sådan bedömning.

Abstract [pt]

As crianças são mais sensíveis à radiação ionizante do que os adultos, devido, entre outros fatores, aos seus órgãos e tecidos estarem em crescimento e à sua maior esperança de vida. Além disso, o tamanho reduzido da cabeça faz com que a glândula tiroideia esteja mais exposta à radiação durante os exames radiológicos dentários. Esta tese incluiu dois grupos de crianças: crianças prematuras e crianças com fenda lábio-palatina.

Anualmente, cerca de 5% das crianças na Suécia nascem prematuras. Esta condição leva frequentemente à necessidade de um maior número de exames médicos hospitalares e, consequentemente, a um maior número de exames radiológicos. No entanto, o contributo da Medicina Dentária para este aumento não é conhecido, um aspeto que foi abordado no primeiro trabalho da tese. Foram analisados 311 diários clínicos de clínicas de Medicina Dentária no Skåne, na Suécia. Os resultados indicaram que as crianças prematuras não apresentavam um estado de saúde oral inferior ao das crianças nascidas a termo, sendo a prevalência de cáries e o número de encaminhamentos para especialistas em Odontopediatria semelhantes.

A fenda lábio-palatina é uma das malformações congénitas da face mais comuns, com uma prevalência de aproximadamente 2 por cada 1000 nascimentos na Suécia. Estas crianças são submetidas a múltiplos tratamentos desde o nascimento. Entre os 9 e os 12 anos, realiza-se o enxerto ósseo alveolar para preencher e encerrar a fenda. Exames radiológicos são necessários tanto antes como após o procedimento cirúrgico, sendo a tomografia computorizada de feixe cónico (CBCT) um dos métodos disponíveis. As imagens de CBCT permitem visualizar a área da fenda, gerar modelos tridimensionais (3D) e estimar a quantidade de osso necessária para o enxerto ósseo alveolar. No segundo trabalho da tese, foram gerados modelos 3D a partir de 53 exames CBCT de fendas unilaterais e bilaterais. Constatou-se, igualmente, que as imagens de CBCT fornecem informação adicional sobre o impacto que a fenda tem nas estruturas anatómicas adjacentes. No entanto, a segmentação manual da área da fenda é um processo moroso.

A inteligência artificial (IA) tem recebido, nos últimos anos, grande atenção, assumindo um papel cada vez mais relevante em diversas áreas da sociedade, nomeadamente na saúde. A Medicina Dentária é uma das áreas em que a IA tem vindo a ser alvo de intensa investigação científica. No terceiro trabalho da tese, foi desenvolvida e testada, com recurso a 88 exames CBCT, uma ferramenta baseada em deep learning. Os resultados mostraram que a ferramenta de IA era capaz de delimitar a área da fenda com elevada precisão e rapidez, bem como estimar a quantidade de osso necessária para o enxerto ósseo alveolar.

A redução das doses de radiação ionizante administradas aos doentes é parte integrante do princípio da otimização, que na União Europeia constitui um requisito legal, com o objetivo de utilizar a menor dose possível de radiação ionizante em cada exame radiológico. O desenvolvimento e teste de protocolos de CBCT de baixa dose são uma estratégia eficaz para reduzir a exposição dos doentes à radiação ionizante. No quarto trabalho da tese, foi testado um protocolo CBCT de baixa dose na avaliação radiográfica realizada após o enxerto ósseo alveolar, em 14 casos de fenda lábio-palatina. Constatou-se que a qualidade de imagem do protocolo CBCT de baixa dose era suficiente para este tipo de avaliação no período pós-operatório.

Place, publisher, year, edition, pages
Malmö: Malmö University Press, 2025. p. 101
Series
Malmö University Odontological Dissertations, ISSN 1650-6065, E-ISSN 2004-9307
Keywords
Preterm birth, Orofacial cleft, Cone-beam computed tomography, Alveolar bone grafting, Artificial intelligence.
National Category
Odontology
Identifiers
urn:nbn:se:mau:diva-79684 (URN)10.24834/isbn.9789178776603 (DOI)978-91-7877-659-7 (ISBN)978-91-7877-660-3 (ISBN)
Public defence
2025-10-24, Auditorium (KL:2370), Faculty of Odontology, Smedjegatan 16, Malmö, 09:15 (English)
Opponent
Supervisors
Note

Paper III in dissertation as manuscript. Not included in the full text online. 

Available from: 2025-09-24 Created: 2025-09-24 Last updated: 2025-12-09Bibliographically approved

Open Access in DiVA

fulltext(1031 kB)1 downloads
File information
File name FULLTEXT01.pdfFile size 1031 kBChecksum SHA-512
2aa080d523e309e78ecb8c32136e536892848405114f4ca83a1969bb354d0a86c22e02d48282b0573edd178c98c2c7d25afd9179f1bce8b5e9f67eecdf6a29fe
Type fulltextMimetype application/pdf

Other links

Publisher's full textPubMedScopus

Authority records

Vicente, AntónioWiedel, Anna-PaulinaBrogårdh-Roth, SusanneHellén-Halme, KristinaShi, Xie-Qi

Search in DiVA

By author/editor
Vicente, AntónioWiedel, Anna-PaulinaBrogårdh-Roth, SusanneHellén-Halme, KristinaShi, Xie-Qi
By organisation
Faculty of Odontology (OD)
In the same journal
Oral surgery, oral medicine, oral pathology and oral radiology
Odontology

Search outside of DiVA

GoogleGoogle Scholar
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

doi
pubmed
urn-nbn

Altmetric score

doi
pubmed
urn-nbn
Total: 28 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf