BAGLS-VF: a comprehensive dataset for glottal area and vocal fold segmentations

Datasets for automatic analysis of laryngeal endoscopic images are scarce, particularly those providing detailed anatomical annotations beyond the glottal area. The BAGLS dataset previously introduced a large collection of laryngoscopic images with expert annotations of the glottal area. To facilitate research on more detailed anatomical modeling and segmentation tasks, we introduce BAGLS-VF, an extension of the BAGLS dataset that includes additional annotations of the left and right vocal folds.

BAGLS-VF contains 59,250 videolaryngoscopic frames with pixel-level annotations of three anatomical structures: glottal area, left vocal fold, and right vocal fold. The dataset is derived from the original BAGLS dataset and was semi-automatically annotated by a trained rater using a standardized annotation protocol. Images originate from high-speed videolaryngoscopic recordings of subjects undergoing clinical examination and cover a wide range of anatomical variability and imaging conditions. In line with the original BAGLS dataset, the images are predominantly grayscale (approximately 97% of the underlying recordings) with a small proportion of RGB recordings (approximately 3%); the segmentation masks are provided as RGB-encoded label images in which the three anatomical regions are stored as separate colour channels. BAGLS-VF is intended to support research in medical image segmentation, laryngeal biomechanics, and computer-assisted diagnosis.