A dataset of laryngeal endoscopic images with comparative study on convolution neural network-based semantic...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Surgical Technology

Background:

• Automated segmentation of anatomical structures is crucial for medical image analysis, supporting autonomous diagnosis and computer/robot-aided interventions.
• Deep convolutional neural networks (CNNs) have surpassed traditional methods but require evaluation in novel environments like transoral endoscopy.
• The human larynx presents unique challenges for segmentation due to its complex anatomy and endoscopic visualization.

Purpose of the Study:

• Evaluate existing deep learning segmentation methods on a new transoral endoscopic dataset of the human larynx.
• Assess the performance of CNN-based semantic segmentation for laryngeal soft tissue.
• Determine the suitability of these methods for clinical applications such as autonomous diagnosis and surgical guidance.

Main Methods:

• Trained four supervised machine learning models (SegNet, UNet, ENet, ErfNet) on a 7-class dataset of 536 manually segmented laryngeal images from two patients.
• Employed data augmentation and network ensembling to enhance segmentation accuracy, using Intersection-over-Union (IoU) as the primary evaluation metric.
• Investigated stochastic inference for model uncertainty and patient-specific fine-tuning for improved transferability.

Main Results:

• A weighted average ensemble of UNet and ErfNet achieved the highest segmentation accuracy, with a mean IoU of 84.7% for laryngeal soft tissue.
• ENet demonstrated the highest computational efficiency, with a mean inference time of 9.22 ms per image.
• Patient-specific fine-tuning with as few as 10 additional images proved sufficient for adapting models to new patients.

Conclusions:

• CNN-based semantic segmentation is effective for endoscopic laryngeal images, offering potential for active constraints and morphological change monitoring.
• The developed segmentation models can aid in autonomous pathology detection and support computer-assisted interventions.
• Future work should focus on larger datasets and self-supervised learning approaches to further enhance segmentation performance.