Image registration using B-splines for contralateral subtraction of chest radiographs

Contralateral subtraction (C-sub) is a computer-aided diagnosis technique for detecting pulmonary nodules in chest radiographs. This technique enhances nodules in a chest image by subtracting its right / left reversed mirror image from the original image. In this paper we propose a C-sub scheme which uses a local registration method newly proposed in this paper. We evaluated the subtraction images obtained by the proposed C-sub scheme using 107 images with nodules in JSRT database, whose subtlety levels are `very subtle', `subtle', and `relatively obvious'. First, the quality of rib elimination in subtraction images was evaluated by a radiologist using a three-point rating of 3 (adequate or better), 2 (poor), and 1 (very poor). As the result, 92.5% cases were scored as 3. Next, a radiologist checked whether or not nodules were clearly depicted at the notified locations in subtraction images. And, the quality of nodule depiction was scored as 3 (clearly depicted), 2 (subtly depicted), or 1 (not depicted). As the result, 73.8% cases were scored as 3 and 7.5% cases were scored as 2. The time needed by the proposed scheme was 19.4 seconds for each image on the average of 107 images by a 2.2GHz Intel PC.