بخش‌بندي تصاوير دهانه حنجره با استفاده از يادگيري عميق

Accurate segmentation of the glottal opening in high-speed videoendoscopy is a fundamental step for quantitative analysis of vocal-fold function an‎d for detecting laryngeal pathologies. Despite recent advances, high heterogeneity across imaging devices, severe illumination variations, an‎d image artifacts in multi-center datasets (e.g., the BAGLS dataset) still make it difficult to obtain both high accuracy an‎d robust performance. In this thesis, we propose several novel deep-learning-based approaches to improve automatic glottis segmentation. First, focusing on input data quality, we introduce a preprocessing method for removing extraneous black borders an‎d a targeted data-augmentation strategy driven by camera-sensitivity analysis. Concurrently, we refine the SA-UNet architecture an‎d selec‎t an appropriate optimization algorithm to enhance extraction of structural features. Second, to exploit the complementary strengths of different architectures, we propose a selec‎tive ensemble learning framework. This framework includes a learned selec‎tor network that, given an input image an‎d several novel auxiliary features, dynamically determines which base model is expected to perform best for that particular image. Third, to reduce false positives in the background, improve model localization, an‎d focus segmentation more precisely on the target region, we develop a two-stage, guided segmentation approach. In this method, an object-detection network (e.g., YOLOv5) is first used to locate a bounding box around the glottal opening. We then introduce a new architecture named BBox-UNet, which accepts the bounding-box information as an auxiliary input by injecting it into the decoder layers an‎d performs the segmentation. Finally, outputs are combined using ensemble methods an‎d a voting scheme. eva‎luation results show that the proposed methods effectively overcome data heterogeneity an‎d significantly improve segmentation accuracy, achieving an Intersection over unio‎n (IoU) of 87.02% an‎d a Dice similarity coefficient (DSC) of 93.06%. These results demonstrate the superiority of the proposed approach over prior methods an‎d indicate its reliability for clinical an‎d decision-support applications.

نادر كريمي

شادرخ سماوي

مينا اميري , فرزانه شايق بروجني