Owing to the recent development of AI technology, various studies on computer-aided diagnosis systems for CT image interpretation are being conducted. In particular, studies on the detection of lung cancer which is leading the death rate are being conducted in image processing and artificial intelligence fields. In this study, to improve the anatomical interpretation ability of CT images, the lung, soft tissue, and bone were set as regions of interest and configured in each channel. The purpose of this study is to select a detector with optimal performance by improving the quality of CT images to detect lung cancer tumors. Considering the dataset construction phase, pixel arrays with Hounsfield units applied to the regions of interest (lung, soft tissue, and bone region) were configured as three-channeled, and a histogram processing the technique was applied to create a dataset with an enhanced contrast. Regarding the deep learning phase, the one-stage detector (RetinaNet) performs deep learning on the dataset created in the previous phase, and the detector with the best performance is used in the CAD system. In the evaluation stage, the original dataset without any processing was used as the reference dataset, and a two-stage detector (Faster R-CNN) was used as the reference detector. Because of the performance evaluation of the developed detector, a sensitivity, precision, and F1-score rates of 94.90%, 96.70%, and 95.56%, respectively, were achieved. The experiment reveals that an image with improved anatomical interpretation ability improves the detection performance of deep learning and human vision.
Research Interests
Biomedical Radar Signal Processing; 생체 레이더 신호 처리; Medical Image Processing; 의료 영상 처리; Embedded System; 임베디드 시스템; Real Time Operating System; 실시간 운영체제