A content-based scheme for CT lung image retrieval

A content based scheme to retrieve computed tomographic images (CT) of the lung is presented. It consists of a visual based user interface to allow the query be made by line drawing the interested (abnormal) regions; and a training scheme to classify the relationship between the images stored in database. The system will output a set of candidate images that are texturally similar to the query image. We marked the abnormal portions of each training image by a polygonal or rectangular shape manually because it requires expert knowledge. Then, the texture features of each marked region are extracted on the DCT or SADCT transform domain. In the training stage, the extracted DCT/SADCT coefficients are fed into a Kohonen self organizing network to find the relationship for classification. In the query stage, the system first checks which texture category the query image is in, then uses some geometrical characteristics to identify the most likely candidate image. Experimental results show that 96% of the queries can be correctly retrieved where the original image is in the candidate set