Banca de DEFESA: Caio Athayde Neves

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : Caio Athayde Neves
DATE: 30/06/2023
TIME: 14:00
LOCAL: Plataforma Teams
TITLE:

Application of deep learning algorithms in the preoperative evaluation of otologic surgery through medical imaging

KEY WORDS:

Otologic Surgical Procedures; Computed Tomography; Deep learning;Temporal bone ; Surgical models

PAGES: 39
BIG AREA: Ciências da Saúde
AREA: Medicina
SUMMARY:

Otologic surgery plays a crucial role in the treatment of hearing loss, infections, and lateral skull base tumors. Accurate segmentation of otologic structures from computed tomography (CT) scans can significantly enhance surgical planning and intraoperative guidance. This manuscript presents two experiments that leverage deep learning algorithms for automated segmentation of key temporal bone structures in CT scans.

In the first experiment, three convolutional neural network (CNN) models (AH-Net, U-Net, ResNet) were compared using 150 manually segmented CT scans. The performance of the models was evaluated based on Dice coefficient, Hausdorff distance, and segmentation speed for the inner ear, ossicles, facial nerve, and sigmoid sinus. AH-Net achieved the best results, with Dice coefficients of 0.91, 0.85, 0.75, and 0.86 for the respective structures. The automated segmentation pipeline demonstrated good correlation with manual segmentation performed by a specialist, opening up possibilities for augmented reality, simulation, and automation applications in otologic procedures.

In the second experiment, a state-of-the-art deep learning algorithm (Swin UNETR) was used to build a prediction model for rapid segmentation of nine key temporal bone structures in 325 clinical CT scans. The model achieved high Dice coefficients, Balanced Accuracy, Volume Similarity, Average Symmetric Surface Distance, and 95th Hausdorff Distance, with a mean processing time of 9.1 seconds per study. This robust model offers valuable new datasets for otologic surgical planning and navigation.

In conclusion, the application of deep learning algorithms in this project demonstrates the potential for improved preoperative evaluation and intraoperative guidance in otologic surgery through automated segmentation of otologic structures from temporal bone CT scans.

 

BANKING MEMBERS:
Presidente - 2292514 - IRUENA MORAES KESSLER
Interno - 3217424 - ANDRE LUIZ LOPES SAMPAIO
Externo à Instituição - HENRIQUE FERNANDES DE OLIVEIRA - HRT
Externo à Instituição - LUCIANA MIWA NITA WATANABE - UNICAMP
Externo à Instituição - MARCIO NAKANISHI - UnB