Feasibility of a deep learning‐based algorithm for automated detection and classification of nasal polyps and inverted papillomas on nasal endoscopic images. Issue 12 (20th June 2021)

• Record Type: Journal Article
• Authors: Girdler, Benton; Moon, Hyun; Bae,, Mi Rye; Ryu, Sung Seok; Bae, Jihye; Yu, Myeong Sang
• Is Part Of: International forum of allergy & rhinology. Volume 11:Issue 12(2021)
• Page Start: 1637
• Page End: 1646
• Publication Date: 2021-06-20
• Subjects: artificial intelligence -- clinical visual assessment -- convolutional neural network -- deep learning -- inverted papilloma -- nasal endoscopy -- nasal polyp; 617.51005
• DOI: 10.1002/alr.22854
• Languages: English
• ISSNs: 2042-6976

Abstract

Abstract: Background: Discrimination of nasal cavity mass lesions is a challenging work requiring extensive experience. A deep learning‐based automated diagnostic system may help clinicians to classify nasal cavity mass lesions. We demonstrated the feasibility of a convolutional neural network (CNN)‐based diagnosis system for automatic detection and classification of nasal polyps (NP) and inverted papillomas (IP). Methods: We developed a CNN‐based algorithm using a transfer learning strategy and trained it on nasal endoscopic images. A total of 99 nasal endoscopic images with normal findings, 98 images with NP, and 100 images with IP were analyzed using the developed CNN. Six otolaryngologists participated in clinical visual assessment. Image‐based classification performance was measured by calculating the accuracy and area under the receiver operating characteristic curve (AUC). The diagnostic performance was compared between the CNN and clinical visual assessment by human experts. Results: The algorithm achieved an overall accuracy of 0.742 ± 0.058 with the following class accuracies: normal, 0.81± 0.14; IP, 0.57 ± 0.07; and NP, 0.83 ± 0.21. The AUC values for normal, IP, and NP were 0.91 ± 0.06, 0.82 ± 0.09, and 0.84 ± 0.06, respectively. The overall accuracy of the CNN model was comparable with the average performance of human experts (0.742 vs. 0.749; p = 0.11). Conclusions: The trained CNN model appears to reliably classify NP and IP of the nasal cavity from nasal endoscopic images; it also yields a reliable reference for diagnosing nasal cavity mass lesions during nasal endoscopy. However, further studies with more test data are warranted to improve the diagnostic accuracy of our CNN model.