Suppression of the Contrast of Ribs in Chest Radiographs by Means of Massive Training Artificial Neural Network

摘要

We developed a method for suppression of the contrast of ribs in chest radiographs by means of a massive training artificial neural network (MTANN). The MTANN is a trainable highly nonlinear filter that can be trained by using input chest radiographs and the corresponding teacher images. We used either the soft-tissue image or the bone image obtained by use of a dual-energy subtraction technique as the teacher image for suppression of ribs in chest radiographs. When the soft-tissue images were used as the teacher images, the MTANN directly produced a "soft-tissue-image-like" image where the contrast of ribs was suppressed. When the bone images were used as the teacher images, the MTANN was able to produce a "bone-image-like" image, and then was subtracted from the corresponding chest radiograph to produce a bone-subtracted image where ribs are suppressed. Thus, the two kinds of rib-suppressed images, i.e., the soft-tissue-image-like image and the bone-subtracted image, could be produced by use of the MTANNs trained with two different teacher images. We applied each of the two trained MTANNs to non-training chest radiographs to investigate the difference between the processed images. The results showed that the contrast of ribs in chest radiographs almost disappeared, and was reduced to less than 10% in both processed images. The contrast of ribs was reduced slightly better in the soft-tissue-image-like images than in the bone-subtracted images, whereas soft-tissue opacities such as lung vessels and nodules were maintained better in the bone-subtracted images. Therefore, the use of the bone images as the teacher images for training the MTANN has produced better rib-suppressed images where soft-tissue opacities were substantially maintained. A method for rib suppression using the MTANN would be useful for radiologists as well as CAD schemes in detection of lung diseases such as nodules in chest radiographs.