Jing Gong1,2, Ji-Yu Liu3, Yao-Jun Jiang4, Xi-Wen Sun3, Bin Zheng5, Sheng-Dong Nie1. 1. School of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, 516 Jun Gong Road, Shanghai, 200093, China. 2. Department of Radiology, Fudan University Shanghai Cancer Center, 270 Dongan Road, Shanghai, 200032, China. 3. Radiology Department, Shanghai Pulmonary Hospital, 507 Zheng Min Road, Shanghai, 200433, China. 4. Department of Radiology, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, 450052, China. 5. School of Electrical and Computer Engineering, University of Oklahoma, Norman, OK, 73019, USA.
Abstract
OBJECTIVES: To develop and test a new multifeature-based computer-aided diagnosis (CADx) scheme of lung cancer by fusing quantitative imaging (QI) features and serum biomarkers to improve CADx performance in classifying between malignant and benign pulmonary nodules. METHODS: First, a dataset involving 173 patients was retrospectively assembled, which includes computed tomography (CT) images and five serum biomarkers extracted from blood samples. Second, a CADx scheme using a four-step-based semiautomatic segmentation method was applied to segment the targeted lung nodules, and compute 78 QI features from each segmented nodule from CT images. Third, two support vector machine (SVM) classifiers were built using QI features and serum biomarkers, respectively. SVM classifiers were trained and tested using the overall dataset with a Relief feature selection method, a synthetic minority oversampling technique and a leave-one-case-out validation method. Finally, to further improve CADx performance, an information-fusion method was used to combine the prediction scores generated by two SVM classifiers. RESULTS: Areas under receiver operating characteristic curves (AUC) generated by QI feature and serum biomarker-based SVMs were 0.81 ± 0.03 and 0.69 ± 0.05, respectively. Using an optimal weighted fusion method to combine prediction scores generated by two SVMs, AUC value significantly increased to 0.85 ± 0.03 (P < 0.05). CONCLUSIONS: This study demonstrates (a) higher CADx performance by using QI features than using the serum biomarkers and (b) feasibility of further improving CADx performance by fusion of QI features and serum biomarkers, which indicates that QI features and serum biomarkers contain the complementary classification information.
OBJECTIVES: To develop and test a new multifeature-based computer-aided diagnosis (CADx) scheme of lung cancer by fusing quantitative imaging (QI) features and serum biomarkers to improve CADx performance in classifying between malignant and benign pulmonary nodules. METHODS: First, a dataset involving 173 patients was retrospectively assembled, which includes computed tomography (CT) images and five serum biomarkers extracted from blood samples. Second, a CADx scheme using a four-step-based semiautomatic segmentation method was applied to segment the targeted lung nodules, and compute 78 QI features from each segmented nodule from CT images. Third, two support vector machine (SVM) classifiers were built using QI features and serum biomarkers, respectively. SVM classifiers were trained and tested using the overall dataset with a Relief feature selection method, a synthetic minority oversampling technique and a leave-one-case-out validation method. Finally, to further improve CADx performance, an information-fusion method was used to combine the prediction scores generated by two SVM classifiers. RESULTS: Areas under receiver operating characteristic curves (AUC) generated by QI feature and serum biomarker-based SVMs were 0.81 ± 0.03 and 0.69 ± 0.05, respectively. Using an optimal weighted fusion method to combine prediction scores generated by two SVMs, AUC value significantly increased to 0.85 ± 0.03 (P < 0.05). CONCLUSIONS: This study demonstrates (a) higher CADx performance by using QI features than using the serum biomarkers and (b) feasibility of further improving CADx performance by fusion of QI features and serum biomarkers, which indicates that QI features and serum biomarkers contain the complementary classification information.