Automatic Detection Of Pulmonary Nodules Based On Ct Images

To research the methods of computer-aided analysis of CT image andcomputer-aided detection(or diagnosis)of lung nodule and their software realization.In the experiments,we selected images of 20 patients underwent CT examination and were diagnosed as lung cancer patients to construct ourexperiment database.There are about 90～120 images per- patient and totally about 2,000images.We only selected 4 images containing suspicious nodules per-patient to constructthe experiment database.In order to test our algorithm,we processed 40 images totallycontaining 42 nodules with varies size.All the images were produced by SIMENSVOLUME ZOOM 4 spiral CT workstation with 512×512 pixel matrix,16-bit allocation,12-bit depth and 4096 gray-level. The pixel size is 0.74 mm×0.74 mm.Firstly,gray-level threshold and region-grow algorithm were applied on the image tosegment lung fields,then 2-Dimensional and 3-Dimensional technologies were used to detect and display the lung nodules.In order to measure some indices of the segrnentednodule,we also performed interactive region-growing technique and artificial neuralnetworks(ANN)classifier on the nodule region.We proposed and realized a lung nodule computer-aided detection systembased on multi-scale nodule enhancement,interactive region-growing segmentation,artificial neural network classifier and 3-dimensional technologies.By using VC++6.0 to develop the software,we realized the 2-Dimensional and 3-Dimensionaldetection and displaying of lung nodules.We processed 40 images containing 42 noduleswith varies size.An experienced radiologist with 10 years experience was asked to detectall the nodules without CAD.And one month later the same radiologist was asked toreview all the images to detect nodules with CAD.ROC analysis was used to assess theperformance of the radiologist,the area of ROC(Az)is 0.96 with CAD,but without CAD,the area is 0.89.We proposed and realized a lung nodule computer-aided detection system.Computer-aided diagnosis is an effective method for the improvements of lung nodule detection performance.