LUNG NODULE SEGMENTATION BASED ON LUNG-RANGE-STANDARDIZATION

Checking radiological image is a very toilsome work for radiologists because it requires long time practice and experienced skill. Therefore, many computer-aided diagnosis (CAD) systems have been introduced to cooperate with radiologists and nowadays many CAD systems based on deep learning exceed human experts in diagnosing accuracy. Nowadays, the much of progress has been made in designing architectures. However, peculiar pre-processing method customized for a certain problem can also increase the model accuracy. After checking the LIDC dataset [44], it has been realized that the locations and sizes of lungs were not regularized. Therefore, in this paper, a new pre- processing method (lung-range-standardization) is proposed in order to improve the general accuracy of lung-related diagnosis systems. And the efficiency of the proposed pre-processing method is validated through comparison between the nodule segmentation model trained using our proposed pre-processing method and the nodule segmentation model, which is trained using the prior pre-processing methods. By using lung-range-standardization we could reduce the difference between train loss and test loss in a great deal (from 0.337 to 0.119).