Prediction of Tumour Site in Larynx Contrast CT Images by Radiomic Feature Analysis

Radiomics, largely used in the field of oncology, uses a variety of complex mathematical computations to extract hundreds of sub-visual quantitative characteristics from radiological medical images. Laryngeal cancer is a frequently observed type of head and neck cancers, known for its unfavorable prognosis. Mineable laryngeal radiomic data provides a goldmine of information waiting to be tapped for better detection, diagnosis and prognosis outcomes. In this paper, we utilized a machine-learning model to classify the laryngeal tumours using the calculated radiomic features present in the CT images. A novel dataset of 303 Head and Neck contrast CT images was collected for the purpose of this study. Slice-wise annotations were created for every tumour by an expert senior radiologist. 3D voxel information, texture features, entropy features, and shape features were extracted for all the images. Feature selection was an integral part for the classification problem as the dimensionality was computationally complex. Sequential forward selection and Logistic regression models were trained to classify the tumour. The work was analyzed with sensitivity, specificity, F1 Score and AUC metrics. Our proposed model developed reached a prediction accuracy of 96% which performed better than other models in the larynx anatomy for prediction of the tumour site.