| Objective To establish a computer-aided detection(CAD) algorithm for pulmonary nodules in CT images, and to evaluate its utility including the preserving effect of the pulmonary nodules and the filtering effect of pulmonary vessels.Material and Methods1. Establishment of the CAD system. Vessels were discriminated from nodules for their morphologic differences, the shape of the former is masslike, while that of the later is sheetlike, by rotating linear mathematical morphological operators with length of7pixels firstly and15pixels secondly. Then vessels were filtered from the thoracic CT images of lung window in JPEG format.2. Evaluation of the CAD system. CT images in JPEG format and lung window of88cases (69cases with5mm slice thickness, and19cases with lmm) with pulmonary nodules were processed by the system. The preserving effect of nodules and filtering effect of vessels were scored. The nodules were grouped by size and type; the lung was classified into3fields by aorta and bottom of the heart. Scores of nodule preserving effect were statistically analyzed with Kruskal Wallis’rank sum test or Mann-Whitney test in different groups to identify whether slice thickness, nodular size or nodular type affect significantly. Scores of vessel filtering effect were statistically analyzed with Kruskal Wallis’rank sum test or Mann-Whitney test in different lung fields, in order to identify whether slice thickness and lung field affect significantly.Result (1) It cost15minutes to process the whole thoracic images with5mm slice thickness, and60minutes with lmm slice thickness.(2) Nodule preserving effect. Information of all the nodules(100%), regardless of nodular size or nodular type, was preserved well in CT images with5mm slice thickness. While information of79%(19/24) nodules was preserved well in CT images with lmm slice thickness.The nodule preserving effect in images with5mm slice thickness was better than that of images with lmm slice thickness in nodules with diameter less than30mm. The nodule preserving effect of images with5mm slice thickness was better than that of images with lmm slice thickness in nodules with cavity or vacuoles. In images with lmm slice thickness, preserving effect of nodules surrounded by pulmonary parenchyma (the third type) was better that of nodules with cavity or vacuoles (the fourth type). No significant differences existed between the other nodular types.(3) Vessel filtering effect.The vessel filtering effect of vessel in images with lmm slice thickness was better than that in5mm slice thickness in the upper and lower lung fields, while no significant differences existed between the images with different slice thickness in middle lung field. Vessels of the three lung fields could be filtered well in images with lmm slice thickness. In images with5mm slice thickness, vessel filtering effect was significant different in different lung fields, the middle field superior to the upper field, the upper field superior to the lower field.Vessel filtering effect in the lower field was poor in some CT images obtained in expiration period,Conclusion Our system makes it that vascular information was filtered well, while the information of nodules was effectively preserved. Vessel filtering effect in images with lmm slice thickness was better than that in images with5mm slice thickness, while nodule preserving effect in images with lmm slice thickness was inferior to that in images with5mm slice thickness. More work is needed to optimize the CAD system, such designing changeable operator length and changeable projection slice numbers. Objective To establish a computer-aided diagnosis system for pulmonary nodules in CT images, and to evaluate the diagnostic performance of radiologists in using the CADx software.Material and Methods1. Establishment of the CADx system.17features of251nodules cases(46benign nodules and205malignant nodules)with pathologic results were identified by2experienced radiologists. The features were translated into values and processed by a weighted negative neighbor algorithm to work out a semi-automatic computer-aided diagnosis system. If features of the new nodule were given, the CADx system will work out the malignant possibility by weighting the similarity of features of the new nodule to that the205malignant and46benign nodules.2. Evaluation of the CADx system.26CT datasets of SPN with pathologic results (14malignant and12benign) were used to assess observers’performance in using the CADx software. Six radiologists differentiate malignant from benign SPN in CT images respectively before and after using CADx software. The sensitivity, specificity and accuracy were compared before and after using CADx software. The CADx software automatically recorded in detail the feature-interpretation of4radiologists. The interpretations in each feature were summarized to determine common features of this serial cases, as well as to describe the consistency of each feature’s interpretation.Results (1) Sensitivity improved apparently from70.24%(59/84) to73.81%(62/84), specificity from59.72%(43/72) to61.11%(44/72), and accuracy from65.38%(102/156) to67.95%(106/156) after using the CADx software.(2) Complete consistency of CAD x recommendation existed in14cases,11of which are correct.(3) Common features of pulmonary nodules include lobulation, spiculated counter, pleural indentation, vacuoles, and spinous sign.(4) The interpretation consistency in the4radiologists:lobular (5/26), spiculated counter (10/26), pleural indentation (9/26), spinous sign (11/26), vacuoles (10/26), vascular aggregation (13/26), mediastinal lymphadenopathy (14/26).Conclusion CADx software designed by our team can slightly improve the diagnosing sensitivity, specificity and accuracy of SPN. More work are needed to improve the system, including effort for increasing accuracy of feature interpretation, open-up system convenient to add new cases and update the training data, more affiliate functions such as automatic retrieval and displaying function of image and pathologic result of cases with several similar features. |
PDF Full Text Request