Author(s)
Noel Phan MD1, Aarthi Ramkrishnan MS 2, Rachel Kominsky MD1, Mark Courey MD1, Li Shen PhD2, Diana N Kirke MBBS MPhil1
Affiliation(s)
1Icahn School of Medicine at Mount Sinai, Department of Otolaryngology – Head & Neck Surgery, New York, NY2Icahn School of Medicine at Mount Sinai, Department of Neuroscience, Shen Laboratory, New York, NY;
Abstract:
Objective: Laryngeal disorders causing dysphonia encompasses a wide range of pathologies. Unilateral vocal fold paralysis (UVFP) is a condition that occurs from dysfunction of the recurrent laryngeal. The purpose of the study is to provide a proof-of-concept that deep convolutional neural networks (DCNNs) can accurately and efficiently differentiate normal laryngeal movement from UVFP.
Methods: Laryngoscopic images were collected by querying the electronic database based on procedural CPT code of laryngoscopy. Patients with ICD-10 codes of UVFP, unilateral vocal cord paralysis, and unilateral vocal cord paresis were identified. Still images were evaluated and only highest quality images were selected to train our DCNNs.
Results: A total of 324 segmented images comprising of 111 UVFP and 213 LPR images were obtained from 324 patients. The DCNN achieved a sensitivity of 0.810 and specificity of 0.906 for detecting UVFP. The final AUC score based on the 5-fold cross-validation was determined to be 0.949 [0.925, 0.971]. Positive predictive value was 0.818 and negative predictive value of 0.901. Of the 111 images of UVFP, 25 samples were misclassified as LPR and of the 213 LPR images, 18 samples were misclassified as having UVFP.
Conclusions: We developed a DCNN algorithm for detecting UVFP using endoscopic images with high sensitivity and high specificity. These preliminary results support further study of DCNNs for clinical detection of abnormal vocal fold movement. The clinically relevant diagnostic ability of deep learning models offers a promising applicability to daily clinical practice where there are shortages of advanced and experienced endoscopists.