An Interpretable Framework Utilizing Radiomics for Laryngeal Cancer Classification Through Narrow Band Imaging

Abstract

Effectively preventing and diagnosing laryngeal cancer is imperative for improving prognosis, particularly in high-risk populations. Early diagnosis in these individuals is crucial, offering timely interventions and higher survival rates. There is a growing demand for computer-assisted early prognosis with transparent and accountable interpretations. Methods and procedures: We present an interpretable classification framework based on radiomics for narrow-band imaging in contact endoscopy for laryngeal cancer. The framework aims to address the challenge of early prognosis with transparency and accountability by offering evidence-supported interpretation. Results: Performance metrics, including accuracy, precision, recall, F1 score, area under the curve (AUC), negative predict value (NPV), balanced accuracy(BA) and diagnostic odds ratio (DOR) were employed to analyze the results. The radiomics-based model demonstrated high recall rate (>94% in 3 out the 5 classifiers employed) and balanced accuracy (>89% in 3 classifiers) in distinguishing between benign and malignant cases. The interpretability analysis shed light on the framework’s behavior evidencing three features as main contributors for classification. Conclusion: Our interpretable classifying framework shows promise as a preliminary clinical diagnostic assistant for laryngeal cancer. Its commitment to transparency and accountability renders it an invaluable resource for conveying insightful information and assessing the framework’s effectiveness. Clinical impact: Our framework addresses the challenge of early diagnosis among high-risk populations, providing a transparent and accountable tool for clinicians and researchers.

Keywords

• Diagnosis
• feature selection
• laryngeal selection
• radiomics