Fig. 1
From: BAGLS, a multihospital Benchmark for Automatic Glottis Segmentation
Workflow for creating the BAGLS dataset. Subjects with varying age, gender and health status were examined at different hospitals with differing equipment (camera, light source, endoscope type). The recorded image data is diverse in terms of resolutions and quality. Next, the glottis was segmented using manual or semi-automatic techniques and the segmentation was crosschecked. The segmented videos were split into a training and a test set. The test set features equal amounts of frames from each hospital. We validated BAGLS by training a deep neural network and found that it provides segmentations closely matching the manual expert segmentations.
