Student Theses, Dissertations and Capstones

Document Type

Thesis - NSU Access Only

Degree Name

Master of Science (M.S.) in Dentistry

Copyright Statement

All rights reserved. This publication is intended for use solely by faculty, students, and staff of Nova Southeastern University. No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, now known or later developed, including but not limited to photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author or the publisher.

Department

College of Dental Medicine

First Advisor

Sundaralingam Premaraj

Date of original Performance / Presentation

January 2023

Publication Date / Copyright Date

1-1-2023

Abstract

Objectives: To evaluate the ability of a deep-learning algorithm to produce automated, error-free segmentation of cleft defects on CBCT images. Background: Cleft lip and palate (CLP) is one of the most common birth defects. Whether associated with a syndrome or occurring as an isolated defect, CLP will have a huge impact on both the affected child and their family. The incidence of orofacial clefts is about 1 in 550 live births. In about 75% of the patients with CLP, the clefting involves the alveolus as well. Bone grafting of alveolar clefts is an essential component in the contemporary management of patients with orofacial clefts. The principal aim of this procedure is to unify the alveolus and create an osseous environment that will support eruption of tooth into the maxillary arch. Outcome assessment is essential in the effective evaluation of clinical management of any illness. For assessment of cleft patient treatment including alveolar bone graft (ABG) procedures, development of methods which could accurately and consistently measure their success or failure are essential and are useful in clinical audit and comparative studies. The clinical audit will drive treatment decisions; comparative studies can impact treatment in a broader sense, as it informs evidence-based education and clinical practice. Cone Beam Computerized Tomography (CBCT) has become the gold standard for visual assessment and quantitative morphometry of cleft lip and palate defects and alveolar bone grafts. Extrapolation of the applicable information provided by CBCT images requires the use of segmentation programs capable of identifying target landmarks and communicating their statistics. Manual segmentation of a CBCT image is labor and time intensive, requiring the segmentation personnel to asses each slice of the image in three dimensions to mark target areas. The evolution of medical image machine learning (ML) techniques has great potential to minimize both the temporal burden and human error. A more advanced version of ML, deep learning (DL), has the ability to recognize and classify image patterns. For the purposes of this study DL will be used as a problem-solving algorithm capable of fully automated detection and quantification of cleft defects. In a previous study for application on cleft defect recognition, segmentation time was decreased from ten hours of manual work to six minutes of algorithm application. Materials and Methods: 30 sets of CBCT images of non-syndromic unilateral cleft palate patients from a private practice in Oklahoma City (Profiles Oral Facial Surgery Experts) were selected. All subjects had received primary lip and palate repair. Each of the 30 subject’s CBCTs reflected a pre-alveolar bone graft status. Subjects with a history of ABG, orthodontic treatment, maxillofacial neoplasia, trauma or orthognathic surgery were excluded. All 30 images underwent manual segmentation. Twenty-four of these 30 manual segmentations served as a training set for the 3D U-Net deep learning algorithm. The trained algorithm was then used to automatically segment a testing sample consisting of the remaining 6 CBCT images. These automatic segmentations are considered predictions, and may need further refinement to more closely approximate the manual segmentation which is considered as the gold standard. Data from all 30 segmented maxilla and defect models were then used for statistical analysis. Volumetric quantification was obtained on ITK-SNAP. Intra- and inter-examiner reliability of manual segmentation were tested using Intraclass Correlation Coefficient (ICC). The accuracy of the automatic segmentation model was evaluated by the Dice similarity coefficient (DSC). Results: The Dice similarity coefficients of the maxilla and defect between manual and automatic segmentation were 0.79  0.02 and 0.61  0.06, respectively. When a DSC value of 1 indicated perfect agreement, these DSC scores reflect reasonable predictive power. An intraclass correlation coefficient (ICC) greater than 0.90 confirms measurement consistency. The ICC to evaluate intra-observer reliability was calculated to be 0.99 for both investigators for both maxilla and defect. The total inter-examiner reliability was defined by an overall ICC of 0.81 with 0.83 and 0.80 for the maxilla and defect, respectively. Conclusion: The 3D U-Net deep learning- based protocol accurately and efficiently segmented the maxilla and defect and shows potential for widespread clinical and large-scale research application.

Disciplines

Dentistry

Keywords

Cleft; Deep Learning; Defect; Dice similarity; U-Net; Volume

Download

Download Full Text (5.2 MB)

Download
To access this thesis/dissertation you must have a valid nova.edu OR mynsu.nova.edu email address and create an account for NSUWorks.

Free My Thesis

If you are the author of this work and would like to grant permission to make it openly accessible to all, please click the Free My Thesis button.

  Contact Author

Share

 
COinS