On Efficient Segmentation of Parapharyngeal Fat Pads From Population-based MRIs: An Obstructive Sleep Apnea Application

The main purpose of our project was to automatically delineate parapharyngeal fat pads from magnetic resonance imaging (MR) data, since these structures are considered important for diagnosis of obstructive sleep apnea syndrome (OSAS). Here, we investigate the problem, discuss possible data choices, compare 2D and 3D networks, and consider several automated processing steps. First, approximately 75% of the sequence slices were excluded from further processing due to a subjects’ anatomy, and it allowed to reduce the training time. Second, 2D and 3D U-Net-based networks were trained on full slices to define the location of fat pads. Although the accuracy values on the validation set were lower in comparison to the human observer, both networks could detect the regions’ location. Finally, the slices were cropped using the ROI information from the previous step and separated in the middle. Another 3D network was trained and the results are comparable to intra-observer variability. Namely, the Dice coefficient on the test set was ≈ 78%, which was comparable to several readings of the same human observer. The presented approach will be applied to big epidemiological data to investigate the effects of the fat pad tissues on OSAS development.