Banca de DEFESA: LAÍS BASTOS PINHEIRO

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : LAÍS BASTOS PINHEIRO
DATE: 07/10/2022
TIME: 15:00
LOCAL: meet.google.com/snh-ctsk-vij
TITLE:

Performance analysis of the Mask R-CNN network for numbering and segmenting teeth in panoramic radiographs

KEY WORDS:

Deep Learning,
Computer vision,
panoramic radiography,
Convolutional Neural Networks,
Supervised Learning.

PAGES: 55
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SPECIALTY: Processamento Gráfico (Graphics)
SUMMARY:

In dentistry, panoramic radiography is an important tool to assist dentists in their diagnoses, monitor oral health, and plan or monitor patient treatment. The large number of indications for use and their advantages of use have boosted studies on the application of deep learning techniques in this type of imaging exam. For the automatic analysis of panoramic radiography, the individual identification of teeth is an indispensable step, since detecting, numbering and segmenting teeth are essential tasks for later stages of automatic diagnosis of these radiographs and generation of automated reports. In this sense, this project proposes to automate the task of identification (detection, numbering and segmentation) of teeth, from the evaluation of neural networks based on deep learning, which delimit, label and segment each tooth detected in the panoramic radiograph. For the evaluation of the overall performance of neural networks, a comprehensive set of panoramic radiographs data with consistent annotations in the image of permanent and deciduous teeth was not found in the literature. To fill this gap, this work contributes an annotated dataset, consisting of the contour of each tooth and manual labeling of each tooth based on the FDI dental notation. This dataset is formatted by 450 images of panoramic radiographs. The notes were taken by dentistry and computer science students, with the help and supervision of professionals in the field of dentistry. From the database created, the proposal of this project is also to evaluate two neural network architectures based on Mask R-CNN: the standard network and another that adds the PointRend module in the segmentation branch. The best performance was achieved with the addition of the PointRend module, which reached 75.3% and 77.3% of mean average precision (mAP), in the numbering and segmentation tasks, respectively, surpassing the standard Mask R-CNN by 1, 2 and 2 percentage points. The aim of the investigation was to find a method that improves segmentation per instance at the limits of teeth, because this is the main obstacle of segmentation methods, which was achieved with the R-CNN Mask plus the PointRend module. It is expected that this study and the new public dataset represent an advance in the automatic processing of image exams in panoramic radiographs, encouraging the proposal of new algorithms to solve the proposed problem.

COMMITTEE MEMBERS:
Presidente - 1914064 - LUCIANO REBOUCAS DE OLIVEIRA
Externo ao Programa - ***.855.755-** - EDUARDO MANUEL DE FREITAS JORGE - UNEB
Externo ao Programa - 6286135 - PAULO SERGIO FLORES CAMPOS - UFBA