Congresso Brasileiro do Sono

Página Inicial » Inscrições Científicas » Trabalhos

Dados do Trabalho


Segmentation of Upper Airways using Deep learning - nnUNet


In the medical field, segmentation is used to identify specific anatomical structures from the rest of the surrounding tissue, such as segmenting tumors in an organ, planning orthognathic surgery, and evaluating structures changes after procedures as orthodontics and surgeries. Computed tomography (CT) has been used as the main imaging modality for the segmentation of anatomical structures, and manual segmentation of the CT scans is the most accurate procedure for upper airway segmentation. However, very time-consuming. Machine learning has a great potential to facilitate this process. Through deep learning (Convolutional Neural Networks-CNN), an algorithm may be developed to automatically segment images or volumes, improving the quality of the segmented structures and decreasing time spent generating the volumes.


Dental sleep medicine is a growing area in which breathing malfunction is a multi-disciplinary issue; the dentist and orthodontist have a great role in it. Screening patients for obstructive sleep apnea and managing oral appliances are dentists' main roles in this multi-disciplinary team. Since CBCT scans are often available at dentists' offices, an automatic tool to evaluate the upper airway (UA) and help to screen the patients for sleep-disordered breathing would help decrease the number of patients underdiagnosed for sleep-disordered breathing problems. The objective of this study is to develop an algorithm using deep learning for segmenting the upper airway (UA).


75 Cone-beam computed tomography (CBCT) obtained for treatment planning were used. All 75 UA geometries were manually segmented; 12000 images from 40 CBCTs were used for training of the convolutional neuro network (CNN), and 10500 images from 35 CBCTs were used for testing its performance. The manually segmented volumes were used as gold standard for comparison with the CNN geometries. The CNN used consisted of the nnU-Net and ResUNet frameworks.


nnUnet showed excellent results with 0.967 dice score, 0.966 Sensitivity, 0.997 Specificity, and 2.45% volume difference compared to the manually segmented geometries. nnUnet took an average of 30 minutes to generate one UA geometry, while manual segmentation took around 5 hours for each geometry.


The use of nnUnet for the segmentation of the UA promotes accurate results in a much faster and automated way.


Upper airway, CBCT, OSA screening, deep learning


Área Clínica


Universidade de Alberta - - Canadá


Silvia Gianoni Capenakas, Alejandro Matos, Manuel Lagravere, Kumaradevan Panithakumar