Hierarchical dynamic convolutional neural network for laryngeal disease classification.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Otolaryngology

Background:

• Laryngeal disease classification presents challenges due to complex anatomical structures and varied imaging viewpoints.
• Existing convolutional neural network methods often overlook sample difficulty variations and exhibit high training complexity.

Purpose of the Study:

• To introduce an end-to-end Hierarchical Dynamic Convolutional Network (HDCNet) for improved laryngeal disease classification.
• To dynamically process input samples based on their classification difficulty, optimizing resource allocation and accuracy.

Main Methods:

• Developed HDCNet, an end-to-end network that processes easy samples with lower resolution and smaller networks, and difficult samples with higher resolution and larger networks.
• Integrated a Feature Reuse Module (FRM) to transfer learned features from smaller to deeper network components, enhancing performance on complex cases.
• Validated the HDCNet on a public laryngeal disease classification dataset.

Main Results:

• HDCNet demonstrated superior performance compared to current state-of-the-art methods in laryngeal disease classification.
• The dynamic processing approach effectively managed varying sample difficulties, leading to improved classification accuracy.
• The Feature Reuse Module contributed to enhanced performance, particularly for intricate laryngeal disease cases.

Conclusions:

• HDCNet offers a more efficient and accurate approach to laryngeal disease classification in medical imaging.
• The dynamic processing strategy addresses limitations of conventional methods by adapting to sample complexity.
• This framework holds promise for advancing automated diagnosis in otolaryngology.