Deep Convolution Neural Network for Laryngeal Cancer Classification on Contact Endoscopy-Narrow Band Imaging

Nazila Esmaeili; Esam Sharaf; Elmer Jeto Gomes Ataide; Alfredo Illanes; Axel Boese; Nikolaos Davaris; Christoph Arens; Nassir Navab; Michael Friebe

doi:10.3390/s21238157

Deep Convolution Neural Network for Laryngeal Cancer Classification on Contact Endoscopy-Narrow Band Imaging

Sensors (Basel). 2021 Dec 6;21(23):8157. doi: 10.3390/s21238157.

Authors

Nazila Esmaeili^{1

2}, Esam Sharaf¹, Elmer Jeto Gomes Ataide^{1

3}, Alfredo Illanes¹, Axel Boese¹, Nikolaos Davaris⁴, Christoph Arens⁵, Nassir Navab², Michael Friebe^{1

6}

Affiliations

¹ INKA-Innovation Laboratory for Image Guided Therapy, Otto-von-Guericke University Magdeburg, 39120 Magdeburg, Germany.
² Chair for Computer Aided Medical Procedures and Augmented Reality, Technical University of Munich, 85748 Munich, Germany.
³ Department of Nuclear Medicine, Medical Faculty, Otto-von-Guericke University Magdeburg, 39120 Magdeburg, Germany.
⁴ Department of Otorhinolaryngology, Head and Neck Surgery, Magdeburg University Hospital, 39120 Magdeburg, Germany.
⁵ Department of Otorhinolaryngology, Head and Neck Surgery, Giessen University Hospital, 35392 Giessen, Germany.
⁶ IDTM GmbH, 45657 Recklinghausen, Germany.

Abstract

(1) Background: Contact Endoscopy (CE) and Narrow Band Imaging (NBI) are optical imaging modalities that can provide enhanced and magnified visualization of the superficial vascular networks in the laryngeal mucosa. The similarity of vascular structures between benign and malignant lesions causes a challenge in the visual assessment of CE-NBI images. The main objective of this study is to use Deep Convolutional Neural Networks (DCNN) for the automatic classification of CE-NBI images into benign and malignant groups with minimal human intervention. (2) Methods: A pretrained Res-Net50 model combined with the cut-off-layer technique was selected as the DCNN architecture. A dataset of 8181 CE-NBI images was used during the fine-tuning process in three experiments where several models were generated and validated. The accuracy, sensitivity, and specificity were calculated as the performance metrics in each validation and testing scenario. (3) Results: Out of a total of 72 trained and tested models in all experiments, Model 5 showed high performance. This model is considerably smaller than the full ResNet50 architecture and achieved the testing accuracy of 0.835 on the unseen data during the last experiment. (4) Conclusion: The proposed fine-tuned ResNet50 model showed a high performance to classify CE-NBI images into the benign and malignant groups and has the potential to be part of an assisted system for automatic laryngeal cancer detection.

Keywords: Deep Convolution Neural Network; cancer; classification; contact endoscopy; larynx; narrow band imaging.

MeSH terms

Endoscopy
Humans
Laryngeal Neoplasms* / diagnostic imaging
Larynx*
Narrow Band Imaging
Neural Networks, Computer