http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
A Classification Method for Magnetic Particle Testing Image Using U-Net
Shunsuke Moritsuka,Tohru Kamiya 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
Magnetic Particle Testing (MPT) is a method for determining the presence or absence of a defect by magnetizing the object to be inspected and sprinkling magnetic powder, which is absorbed by the defective part such as a crack and appears as a magnetic powder pattern, which is then evaluated by a specialist. By using the MTP, inspection can be performed without breaking the object to be inspected. However, there are some problems such as the possibility of overlooking defects. In this paper, to solve the problems we develop a classification method of defect images by deep learning for the automation of MPT. The proposed method is based on the structure of U-Net, which has excellent segmentation capability in image processing, and performs segmentation using an improved model that adds convolutional layers to U-Net. Then, an algorithm that combines the result with the last part of the encoder of U-Net is used to discriminate the presence or absence of defects. Using this method, defects were classified from the images obtained during MPT. The results showed that Accuracy of 85.8%, TPR of 65.2%, and FPR of 13.8%.
Extraction of Cervical Lymph Nodes using Improved U-Net++
Nozomi Shime,Tohru Kamiya,Takayuki Ishida 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Early detection and treatment of the lymph node are important since swelling of the neck is a likely factor in systemic metastasis of cancer. One of the diagnoses of cervical swelling is a CT scan, which has a beneficial influence on the diagnosis of the disease. However, the reading of CT images is burdened by the large number of images, which increases the physicians workload. In addition, since it is based on the subjective judgment of the physician, there may be discrepancies in diagnostic results and undetected cases due to differences in experience. A means of solving these problems requires a CAD system that provides a second opinion to the physician. Therefore, this paper proposes a segmentation method of cervical lymph node region for the purpose of developing a CAD system for the diagnosis of cervical lymphadenopathy from CT images. The proposal method is a CNN model with U-Net++ as the backbone, introducing CBAM (convolution block attention module) and dual-branch multi-scale attention module. The proposed method was applied to CT images of 11 cases, yielding IoU of 62.29, confirming its validity.
Kosei Tamura,Tohru Kamiya,Masafumi Oda,Yasuhiro Morimoto 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Root resorption is a pathological condition which is characterized by the loss of the tooth root. Root resorption is not painful in its early stages. As a result, many people who are potentially affected and the condition are often left untreated until it is detected during regular check-ups. If detected early, good treatment results can be achieved, whereas failure to treat the condition properly can lead to tooth extraction. However, the root resorption is currently difficult to detect on panoramic radiographs and may be treated as caries after it becomes painful. The aim of this paper is to identify root resorption from panoramic X-ray images using a deep metric learning algorithm. As a loss function for distance learning, it is known that the loss function in angle space is consistent. Therefore, a loss function is defined and trained using the cosine value of the angle between the feature and the center position to improve the discrimination performance. We obtained experimental results based on 150 image sets with 0.80 of accuracy, 0.62 of TPR, 0.19 of FPR and 0.78 of AUC, respectively.
Detection of Driver Gene Mutations from Thoracic CT Images Based on LightGBM with Radiomics Features
Shion Watanabe,Tohru Kamiya,Takashi Terasawa,Takatoshi Aoki 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Lung cancer is one of the most common cancers worldwide and has become a general medical problem. To lessen the risk of death, early detection and treatment is particularly required. The patients can use molecularly targeted drugs when the driver gene mutations of the cancer are detected, but invasive biopsies are required. So, development of new methods to detect it noninvasively and in a short time are expected. we propose a new machine learning method for identifying the presence or absence of driver gene mutations of lung cancer on Thoracic CT Images that is a non-invasive, in a short time, and low-cost CAD (Computer Aided Diagnosis) system. In the proposed method, radiomics features are given as explanatory variables in addition to Thoracic CT Images, and supervised learning using LightGBM is performed to conduct binary classification with/without driver gene mutations.
Identification of abnormal tissue from CT images using improved ResNet34
Naoya Honda,Tohru Kamiya,Shoji Kido 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
In recent years, CT examinations have been widely used as a screening method to detect lung cancer. However, reading enormous CT images become a heavy burden to the physician. To avoid this problem, computer-aided diagnosis systems have been introduced on CT screening. In general, physicians consider patient information in addition to image information when they make a diagnosis, new efforts are being made to improve the accuracy of diagnosis by mimicking this information with a machine. In this paper, we propose a method for identifying pulmonary nodules by adding medical record information to images to improve the accuracy of diagnosis. We classify nodules from unknown data by assigning branching information of vascular opacities, straight vascular shadows, and nodular shadows as labeled image, which are a cause of misrecognition based on image features in machine learning. In the experiment, the classification accuracy of the nodule class was improved by adding clinical information to 644 images including 161 nodal images.