|
Extraction of Lung Tumor Regions from Thoracic CT Images Using An Improved U-Net |
|
Reo Takahashi, Tohru Kamiya(Kyushu Institute of Technology, Japan), Takashi Terasawa, Takatoshi Aoki(University of Occupational and Environmental Health, Japan) |
| We propose a deep learning model for automatic extraction of lung tumor regions from chest CT images to reduce radiologists' workload and enhance early detection of lung cancer. Our model is based on U-Net and incorporates CBAM and MultiRes Blocks. Experimental results show a 4.2% improvement in Dice and a 4.6% improvement in IoU compared to the original U-Net. This could potentially help in the early diagnosis and treatment of lung cancer by making the screening process more efficient and reducing human error. |
|
|
A Method for Denoising from Low-Dose CT images Based on Noise2Noise |
|
Shuji Sawada, Tohru Kamiya(Kyushu Institute of Technology, Japan), Guangxu Li(Tiangong University, China), Seiichi Murakami(Junshin Gakuen University, Japan) |
| Japan has the highest number of CT scanners per million people among developed countries. Screening of high-risk populations by means of low-dose CT has been verified to be able to reduce lung cancer mortality. However, the low-dose CT image noise should be accurately estimated in order to achieve good image quality as normal dose CT. In this paper, we apply a denoising method using convolutional neural networks without use of the clean images during learning phase. The CNN model is an improvement on the model that achieved high performance in the super-resolution task. Comparison with normal dose CT images and evaluation of image quality using PSNR showed that the proposed method is useful. |
|
|
Automatic Segmentation of Finger Bone Regions from CR Image Using An Improved HRNet |
|
Takumu Hiraoka, Tohru Kamiya(Kyushu Institute of Technology, Japan), Takatoshi Aoki(University of Occupational and Environmental Health, Japan) |
| In Japan, the increase in the number of people requiring nursing care due to the super-aging society has become a serious problem. One of the main causes of people identified as requiring nursing care is bone disease. However, the diagnosis of this disease by physicians has the problem that the diagnosis lacks objectivity and reproducibility. To solve this problem, computer-aided diagnosis (CAD) systems are being developed. Therefore, this paper proposes a segmentation method of phalanx region from CR images for the purpose of developing a CAD system. The proposed method is HRNet + JPU + U-Net. Experimental results on 101 cases confirmed the usefulness of the proposed method. |
|
|
Facial Symmetry Analysis for Cleft Lip Patients from 4D Point Cloud Data |
|
Narumi Kihara(Kyushu Institute of Technology, Japan), Namiko Kimura-Nomoto(Kagoshima University, Japan), Takako Okawachi(National Hospital Organization Kagoshima Medical Center, Japan), Norifumi Nakamura(Kagoshima University, Japan), Guangxu Li(Tiangong University, Japan), Tohru Kamiya(Kyushu Institute of Technology, China) |
| Cleft lip is one of the most common birth defects. Several operations are performed to form a natural lip. A problem with these operations is that the criteria for the facial symmetry are unclear. Based on this background, we propose a method for evaluating the facial symmetry using 4D point cloud data. We evaluate the facial symmetry in two ways. One is based on the temporal changes in the face landmarks, and the other is based on the center of gravity. The experiment is performed using artificially generated 4D data. As an experimental result, it is shown that our method can find point correspondences with smaller error than comparative methods. |
|