Classification of Annotated Pulmonary Nodules with Pathologically Confirmed Malignant, Benign and Metastasis Cases

Abstract

This paper presents a novel framework for combining well known shape, texture, size and resolution informatics descriptor of solitary pulmonary nodules (SPNs) detected by computed tomography (CT). The proposed methodology evaluates the performance of classifiers in differentiating benign, malignant as well as metastasis SPNs with 246 Chest CT scan of patients. 489 unique nodules were extracted from these CT scans. Using available 80 pathologically confirmed cases, 211 nodules were labeled as malignant (M), benign (B) and metastasis (MT). Performance evaluation was carried out using 83 extracted features of the nodules. A reduced set using principal component analysis (PCA) resulting 12 features was also considered. Support vector machine (SVM), decision tree and naive Bayes classifiers were evaluated for two classes (M and B) as well as for three classes (M, B and MT). The evaluation results indicate that most of the selected features have important contribution in differentiating SPNs. The Receiver –Operator- Characteristic curve (ROC) is plotted for two classes as well as for three class classification to strike a balance between sensitivity and specificity. Higher value of specificity obtained indicates the potential of the methodology to avoid unnecessary biopsies. These results can be used to build a highly efficient feature index of a content-based medical image retrieval system with pulmonary nodules system for CT.