Inverted papilloma and nasal polyp classification using a deep convolutional network integrated with an attention...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Otolaryngology

Background:

• Inverted papilloma (IP) is a sinus neoplasm with malignant potential, while nasal polyps (NP) are common sinus masses.
• Distinguishing IP from NP on computed tomography (CT) is challenging for otolaryngologists due to subtle visual differences.
• Accurate preoperative classification of IP and NP is crucial for effective treatment and clinical management.

Purpose of the Study:

• To develop and evaluate a neural network model for classifying inverted papilloma (IP) and nasal polyps (NP) using CT images.
• To analyze the model's ability to discriminate between these two common sinonasal conditions.

Main Methods:

• A deep convolutional neural network (CNN) incorporating attention mechanisms (SE-DenseNet) was utilized for classification.
• The model combines DenseNet for feature extraction and SENet for channel attention to enhance feature discriminativeness.
• Interpretability was assessed by analyzing heatmaps of the final convolutional layer.

Main Results:

• The SE-DenseNet model achieved an average accuracy (Acc) of 88.4% and an area under the curve (AUC) of 0.87 on a dataset of 3382 CT slices from 136 patients.
• Visualized heatmaps demonstrated the model's capability to effectively locate sinonasal lesions.
• The model successfully differentiated between IP and NP.

Conclusions:

• The proposed SE-DenseNet model offers an accurate and efficient tool for otolaryngologists to diagnose IP and NP in CT scans.
• The model's reliability is supported by its performance metrics and visualized heatmaps.
• This AI-driven approach can aid in preoperative recognition and clinical decision-making for sinonasal conditions.