Keywords: Narrow Band Imaging, Laryngeal Cancer, Cross Entropy Loss, Supervised Contrastive Loss
TL;DR: This paper proposes a method to learn robust representations for laryngeal cancer classification from narrow band images by using wavelet scattering features in addition to deep features.
Abstract: Narrow Band Imaging (NBI) is increasingly being used in laryngology because it increases the visibility of mucosal vascular patterns which serve as important visual markers to detect premalignant, dysplastic, and malignant lesions. To this end, deep learning methods have been used to automatically detect and classify the lesions from NBI endoscopic videos. However, the heterogeneity of the lesions, illumination changes due to phlegm on the mucosa, and imaging artifacts such as blurriness make inter-patient endoscopic videos exhibit diverging image distributions. Therefore, learning representations that are robust to image distribution changes can be beneficial and improve the generalizing capability of the convolutional neural network (CNN). To this end, we propose a dual branch CNN that learns robust representations by combining deep narrow band features and wavelet scattering transform features of the narrow band images to classify vocal cord NBI images into malignant and benign classes. We show the generalizing capability of our learnt representation by training our neural network using two different losses: cross-entropy (CE) loss and supervised contrastive (SupCon) loss.
Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: novel methodological ideas without extensive validation
Primary Subject Area: Application: Endoscopy
Secondary Subject Area: Detection and Diagnosis
Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.