Re-Training of Convolutional Neural Networks for Glottis Segmentation in Endoscopic High-Speed Videos

Area of Science:

• Medical Imaging
• Computational Biology
• Laryngology

Background:

• Endoscopic high-speed video (HSV) systems are crucial for assessing vocal fold dynamics.
• Advancements in HSV technology necessitate updates to neural network (NN) models for accurate image processing.
• Existing NN models require re-training to accommodate new recording modalities and data variations.

Purpose of the Study:

• To evaluate re-training strategies for convolutional neural networks (CNNs) used in HSV image segmentation.
• To improve the accuracy of NN-based image processing for diverse HSV systems.
• To investigate the impact of data diversity and re-training on segmentation performance.

Main Methods:

• Developed a new dataset (BAGLS-RT) with 21,050 images from varied HSV systems, light sources, and resolutions.
• Re-trained a baseline CNN (trained on 58,750 images) using the expanded dataset.
• Compared segmentation accuracy (mIoU) across different re-training approaches, including fine-tuning with dynamic knowledge distillation.

Main Results:

• Increasing data diversity via preprocessing improved segmentation accuracy by 6.35% (mIoU).
• Subsequent re-training further enhanced segmentation performance by an additional 2.81% (mIoU).
• Fine-tuning with dynamic knowledge distillation yielded the most effective re-training results.

Conclusions:

• Data augmentation and re-training are effective for boosting HSV image segmentation quality.
• Dynamic knowledge distillation is a promising re-training method for adapting CNNs to new data.
• Catastrophic forgetting remains a challenge during re-training, requiring careful consideration to retain previously learned knowledge.