Automatic detection and classification of nasopharyngeal carcinoma on PET/CT with support vector machine

Abstract

Positron emission tomography/computed tomography (PET/CT) has established values for imaging of head and neck cancers, including the nasopharyngeal carcinoma (NPC), utilizing both morphologic and functional information. In this paper, we introduce a computerized system for automatic detection of NPC, targeting both the primary tumor and regional nodal metastasis, on PET/CT. Candidate lesions were extracted based on the features from both PET and CT images and a priori knowledge of anatomical features and subsequently classified by a support vector machine algorithm. The system was validated with 25 PET/CT examinations from 10 patients suffering from NPC. Lesions manually contoured by experienced radiologists were used as the gold standard. Results showed that the system successfully identified all 53 hypermetabolic lesions larger than 1cm in size and excluded normal physiological uptake in brown fat, muscles, bone marrow, brain, and salivary glands. The system combined both imaging features and a priori clinical knowledge for classification between pathological and physiological uptake. Preliminary results showed that the system was highly accurate and promising for adoption in clinical use.