Improvement diagnostic accuracy of sinusitis recognition in paranasal sinus X-ray using multiple deep learning models

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Paranasal sinus (PNS) X-ray imaging, particularly Waters' view, is utilized for diagnosing sinusitis based on opacification or air/fluid levels.
• Maxillary sinusitis detection remains a key area for diagnostic improvement in radiology.

Purpose of the Study:

• To assess the feasibility of identifying maxillary sinusitis features on PNS X-ray images.
• To develop an effective consensus method using multiple deep learning models for sinusitis classification.

Main Methods:

• A dataset of 4,860 PNS X-ray scans (2,430 normal, 2,430 maxillary sinusitis) was used, split into training, validation, and test sets.
• Three convolutional neural network (CNN) models (VGG-16, VGG-19, ResNet-101) were employed.
• A majority decision algorithm integrated these CNNs to establish a consensus diagnosis.

Main Results:

• The majority decision algorithm achieved the highest accuracy (94.1% internal, 94.12% external test datasets).
• Individual CNN models showed strong performance: VGG-16 (87.4% internal, 87.58% external), VGG-19 (90.8% internal, 87.58% external), and ResNet-101 (93.7% internal, 92.12% external).
• The majority decision algorithm demonstrated superior lesion detection capabilities.

Conclusions:

• The majority decision algorithm significantly outperformed individual CNN models in accuracy and lesion detection for maxillary sinusitis.
• Deep learning applied to PNS X-ray images provides a valuable adjunct for classifying maxillary sinusitis.