Detecting oropharyngeal carcinoma using multispectral, narrow‐band imaging and machine learning

OBJECTIVE: To determine if multispectral narrow-band imaging (mNBI) can be used for automated, quantitative detection of oropharyngeal carcinoma (OPC). STUDY DESIGN: Prospective cohort study. METHODS: Multispectral narrow-band imaging and white light endoscopy (WLE) were used to examine the lymphoepithelial tissues of the oropharynx in a preliminary cohort of 30 patients (20 with biopsy-proven OPC, 10 healthy). Low-level image features from five patients were then extracted to train naïve Bayesian classifiers for healthy and malignant tissue. RESULTS: Tumors were classified by color features with 65.9% accuracy, 66.8% sensitivity, and 64.9% specificity under mNBI. In contrast, tumors were classified with 52.3% accuracy (P = 0.0108), 44.8% sensitivity (P = 0.0793), and 59.9% specificity (P = 0.312) under WLE. Receiver operating characteristic analysis yielded areas under the curve (AUC) of 72.3% and 54.6% for classification under mNBI and WLE, respectively (P = 0.00168). For classification by both color and texture features, AUC under mNBI increased (80.1%, P = 0.00230) but did not improve under WLE (below 55% for both models, P = 0.180). Cross-validation with five folds yielded an AUC above 80% for both mNBI models and below 55% for both WLE models (P = 0.0000410 and 0.000116). CONCLUSION: Compared to WLE, mNBI significantly enhanced the performance of a naïve Bayesian classifier trained on low-level image features of oropharyngeal mucosa. These findings suggest that automated clinical detection of OPC might be used to enhance surgical vision, improve early diagnosis, and allow for high-throughput screening. LEVEL OF EVIDENCE: NA. Laryngoscope, 2514-2520, 2018.