کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Robust Glottis Segmentation by Selective Image Augmentations

نویسندگان :

Amir Hosein Kamali¹ Behzad Mirmahboub² Nader Karimi³ Shadrokh Samavi⁴

1- دانشگاه صنعتی اصفهان 2- دانشگاه صنعتی اصفهان 3- دانشگاه صنعتی اصفهان 4- دانشگاه سیاتل

کلمات کلیدی :

Glottis Segmentation،Multi-Camera Dataset،Selective Augmentation

چکیده :

Automated segmentation of the glottis from laryngeal videoendoscopy is crucial for the assessment of vocal fold function and the diagnosis of voice disorders. While deep learning models, particularly U-Net architectures, have shown promise, their performance on the diverse and challenging Benchmark for Automatic Glottis Segmentation (BAGLS) dataset leaves room for improvement. Because images in this dataset are captured using various cameras, we investigate the effect of image augmentation for a subset of cameras. Our experiments with a robust encoder–decoder network architecture show that proper selection of the image subset for augmentation during the training phase improves segmentation results. We also apply test-time augmentation for further improvement. Our method achieves a Dice Similarity Coefficient (DSC) of 90.98% and an Intersection over Union (IoU) of 83.64% on the BAGLS test set, which substantially outperforms previous state-of-the-art models and establishes a new performance benchmark for this task.