Image Retrieval and Identification of Abdominal Organs based on Convolutional Features

Abstract

Medical imaging contributes to the field of medicine, including diagnosis, treatment, clinical decisionmaking, disease research, and teaching. As the technology of medical imaging rapidly advances, the amount of relevant data has increased rapidly. To efficiently manage and retrieve large amounts of data, content-based medical image retrieval (CBMIR) system is required. Among various medical images, abdominal computed tomography (CT) images are often used clinically. However, if the user does not specify the organ to be retrieved and only inputs the abdominal CT image, the retrieval cannot be performed as many organs exist in a CT image. Therefore, we established a content-based medical image retrieval system to quickly and accurately retrieve similar images according to the user-specified organ as the query image. For the user to indicate a specific organ to query, we proposed inputting the query CT image with the organs delineated and with the target image database. The medical images used in this study were all abdominal CT images provided by Mackay Memorial Hospital, including 282 gallbladders, 424 left kidneys, 362 right kidneys, 638 livers, 442 spleens, 450 stomachs, and 229 pancreases, a total of2827 images. A medical expert marked the CT images with the organs delineated and established the corresponding individual organ datasets. In the experiment, a deep convolutional neural network was used to establish a classification model and extract the features from all images in the target image database. When querying, the system automatically calculated the cosine similarity between the features of the query image and the features of all images in the target image database. The first ten images with the highest similarity were used as the retrieval result. Finally, the mean average precision (mAP) for every organ was used to evaluate the retrieval performance of the system. The experimental results showed that the mAP of the medical image retrieval system proposed in this study was 0.94. this indicates that the system effectively retrieved the abdominal CT images of the query-specific organs.