Deep Learning-Based Detection of Glottis Segmentation Failures.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate medical image segmentation is essential for clinical applications, yet noise and variability pose significant challenges.
• Precise glottis segmentation from high-speed videos is critical for voice research and diagnostics.
• Manual identification of segmentation failures is time-consuming and inefficient.

Purpose of the Study:

• To develop and evaluate the first deep learning framework for automated detection of faulty glottis segmentations.
• To improve the efficiency and reliability of glottis segmentation analysis in medical imaging.
• To establish a robust method for identifying suboptimal segmentation outcomes.

Main Methods:

• Generated faulty glottis segmentations using a poorly performing neural network and a novel knowledge-driven perturbation procedure on public datasets.
• Applied extensive data augmentation and image transformations to create diverse failure cases.
• Trained a ResNet18 neural network with custom loss functions to predict segmentation quality scores (IoU).

Main Results:

• The proposed deep learning model achieved 88.27% classification accuracy in detecting faulty segmentations.
• The system demonstrated high specificity (91.54%) in identifying incorrect segmentations.
• The method effectively predicts Intersection over Union (IoU) scores, enabling classification via a fixed threshold.

Conclusions:

• The developed deep learning approach offers an effective and automated solution for detecting faulty glottis segmentations.
• This framework significantly reduces the need for manual review, streamlining voice research and diagnostic processes.
• The study highlights the potential of AI in enhancing the quality control of medical image segmentation.