Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Linear and non‑linear methods for the prediction of mandible morphology
Terblanche, J. 2025. Linear and non‑linear methods for the prediction of mandible morphology. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/72000e8b-9b05-40d8-91c1-361327b19916
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stelenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613918435737600 |
|---|---|
| access_status_str | Open Access |
| author | Terblanche, Jacques |
| author2 | Van der Merwe, Johan |
| author_browse | Terblanche, Jacques Van der Merwe, Johan |
| author_facet | Van der Merwe, Johan Terblanche, Jacques |
| author_sort | Terblanche, Jacques |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Terblanche, J. 2025. Linear and non‑linear methods for the prediction of mandible morphology. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/72000e8b-9b05-40d8-91c1-361327b19916 |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/132406 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:43:46.817Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stelenbosch : Stellenbosch University |
| publisherStr | Stelenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/132406 Linear and non‑linear methods for the prediction of mandible morphology Terblanche, Jacques Van der Merwe, Johan Laubscher, Ryno Stellenbosch University. Faculty of Engineering. Dept. of Mechanical and Mechatronic Engineering. Mandible -- Morphology Cephalometry Morphogenesis Cranial manipulation UCTD Terblanche, J. 2025. Linear and non‑linear methods for the prediction of mandible morphology. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/72000e8b-9b05-40d8-91c1-361327b19916 Thesis (MEng)--Stellenbosch University, 2025. ENGLISH ABSTRACT: Patient-specific methods that provide repeatable, surgeon-independent diagnoses for assessing facial harmony are essential in orthognathic surgery. Current model-based approaches typically assume a linear relationship between facial geometry, which may not accurately capture the complex interactions between upper and lower facial structures. No studies have compared linear and non-linear models for estimating lower facial cephalometric measurements based on upper facial cephalometric measurements. Therefore, this project investigates the efficacy of both linear and non-linear models, specifically ordinary least squares regression (OLS), random forest (RF), multilayer perceptron (MLP), and mixture density networks (MDN), to estimate lower facial measurements from upper facial data. Over 600 computed tomography scans were processed, of which 155 were further landmarked to extract upper and lower facial measurements. The models were implemented using Scikit-learn and PyTorch frameworks, with hyperparameter optimisation conducted via grid search and Bayesian search. In evaluating the models, the test set mean absolute error (MAE) for distance measurements were 2.77mm (MLP), 2.79mm (MDN), 2.95mm (OLS), and 2.91mm (RF), while for angle-based measurements, the MAEs were 3.09°, 3.11°, 3.07°, and 3.12°, respectively. All models demonstrated comparable performance, with neural network-based methods, specifically the MLP, having marginally more predictions below clinical tolerances. Notably, under more lenient thresholds, the mixture density network (MDN) is advantageous due to offering comparable performance to the MLP and the ability to estimate uncertainty. Consequently, while non-linear methods such as the MLP and MDN offer marginal improvements in prediction performance, additional refinement is necessary across all models to attain clinical applicability. AFRIKAANSE OPSOMMING: Pasiëntspesifieke metodes wat herhaalbare, chirurg-onafhanklike diagnoses vir gesigharmonie-assessering verskaf, is noodsaaklik in ortognatiese chirurgie. Huidige modelgebaseerde benaderings veronderstel tipies ’n lineêre verband tussen gesigsmeetkunde, wat moontlik nie die komplekse interaksies tussen boonste en onderste gesigstrukture akkuraat vasvang nie. Vorige studies het nog nie lineêre modelle met nie-lineêre modelle vergelyk vir die skatting van onderste kefalometriese gesigsmetings uit boonste kefalometriese gesigsmetings nie. Daarom ondersoek hierdie projek die doeltreffendheid van beide lineêre en nie-lineêre modelle, spesifiek gewone kleinste kwadraat-regressie (GKK), ewekansige woud (EW), meerlaagperseptron (MLP) en mengseldigtheidsnetwerke (MDN), om laer gesigsmetings uit boonste gesigdata te skat. Meer as 600 rekenaartomografie-skanderings is verwerk, wat gelei het tot 155 skanderings wat verder gemerk is om boonste en onderste gesigsmetings te onttrek. Die modelle is geïmplementeer met behulp van Scikit-learn- en PyTorch-raamwerke, en model-hiperparameteroptimalisering is uitgevoer via roostersoektog en Bayesiaanse soektog. Die toetsstel se gemiddelde absolute fout (MAE) vir afstandmetings was 2.77mm (MLP), 2.79mm (MDN), 2.95mm (GKK) en 2.91mm (EW), terwyl die MAE’s vir hoekgebaseerde afmetings onderskeidelik 3.09°, 3.11°, 3.07° en 3.12° was. Alle modelle het vergelykbare prestasie getoon, met neurale netwerkgebaseerde metodes, spesifiek die MLP, met marginaal meer voorspellings wat aan kliniese toleransies voldoen. Opmerklik is dat onder meer toegeeflike drempels, die mengseldigtheidsnetwerk voordele bied as gevolg van sy soortgelyke prestasie as die MLP en sy vermoë om onsekerheid te skat. Gevolglik, terwyl nie-lineêre metodes soos die MLP en MDN effense verbeterings in voorspellingsprestasie bied, is bykomende verfyning oor alle modelle nodig om kliniese toepaslikheid te bereik. Masters 2025-06-06T07:19:49Z 2025-06-06T07:19:49Z 2025-03 Thesis https://scholar.sun.ac.za/handle/10019.1/132406 en Stellenbosch University xvi, 102 pages : illustrations application/pdf Stelenbosch : Stellenbosch University |
| spellingShingle | Mandible -- Morphology Cephalometry Morphogenesis Cranial manipulation UCTD Terblanche, Jacques Linear and non‑linear methods for the prediction of mandible morphology |
| title | Linear and non‑linear methods for the prediction of mandible morphology |
| title_full | Linear and non‑linear methods for the prediction of mandible morphology |
| title_fullStr | Linear and non‑linear methods for the prediction of mandible morphology |
| title_full_unstemmed | Linear and non‑linear methods for the prediction of mandible morphology |
| title_short | Linear and non‑linear methods for the prediction of mandible morphology |
| title_sort | linear and non linear methods for the prediction of mandible morphology |
| topic | Mandible -- Morphology Cephalometry Morphogenesis Cranial manipulation UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/132406 |
| work_keys_str_mv | AT terblanchejacques linearandnonlinearmethodsforthepredictionofmandiblemorphology |