Vocal cord lesions classification based on deep convolutional neural network and transfer learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Otolaryngology

Background:

• Laryngoscopy is the primary method for diagnosing vocal cord lesions (VCLs).
• Current diagnosis relies on subjective visual inspection by otolaryngologists.
• There is a need for more objective diagnostic tools.

Purpose of the Study:

• To develop an objective computer-aided diagnosis system for VCLs.
• To utilize deep convolutional neural networks (DCNNs) and transfer learning for VCL classification.
• To improve the accuracy and efficiency of VCL diagnosis.

Main Methods:

• A DCNN backbone combined with transfer learning was finetuned on a laryngoscopy image dataset.
• The system was trained on a collected database of laryngoscopy images.
• Performance was evaluated against other DCNN models using F1 scores and receiver operating characteristic curves.

Main Results:

• The proposed system achieved 80.23% accuracy, 0.7836 F1 score, and 0.9557 AUC for four VCL classes.
• It demonstrated high accuracy (0.939) and AUC (0.9828) for urgent and non-urgent lesion detection.
• The system outperformed clinicians in classifying certain VCLs with minimal time cost.

Conclusions:

• The DCNN-based system effectively distinguishes common VCLs and normal cases.
• It holds practical potential for improving diagnostic efficacy in VCL examinations.
• The system can serve as an objective auxiliary tool integrated into laryngoscopy workflows.