Fine-grained Classification of Bone Scintigrams by Using Radiomics Features

T o develop fine-grained diagnosis of lung cancer subtypes (i.e., adenocarcinoma, squamous cell carcinoma, small cell carcinoma) in regional bone scans, in this work, we propose a radimocis-based method. In the feature selection stage, multiple feature selection methods are used to extract key features from bone scintigrams. Several classifiers are developed using random forest, decision tree, naive Bayes and other algorithms to classify lesions. Experimental evaluation conducted on clinical SPECT images shows that radimocis features including morphological features and higher-order features have a significant effect on the classification performance. The best classification performance is achieved by the proposed model using LASSO and random forest algorithm, obtaining scores of 0.6202, 0.6253, 0.6202, and 0.6163 for accuracy, precision, recall, and F-1 scores, respectively.