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.