|
Remote Sensing Image Registration Based on Improved Geometric-matching CNN |
|
Futa Morishima, Huimin Lu, Tohru Kamiya(Kyushu Institute of Technology, Japan) |
| Remote sensing image registration is an important pre-processing step in detection of environmental changes. Currently, image registration methods based on deep learning are gaining attention. The increase in image size leads to higher computational costs during training and estimation of deep learning models. Then, we propose a method that reduces the number of parameters of the model to lower the computational cost while maintaining the accuracy. We modified GMCNN by adding CSA module, SE layer and point-wise convolution layer. The improved GMCNN decreases the grid MSE by 0.0037 compared to the conventional GMCNN. It also reduces the number of parameters by 49.6%.
|
|
|
Identification of Root resorption on Panoramic X-ray Image Based on EfficientNet |
|
Keisuke Sakata, Tohru Kamiya(Kyushu Institute of Technology, Japan), Masafumi Oda, Yasuhiro Morimoto(Kyushu Dental University, Japan) |
| This paper proposes an image analysis method to identify the presence or absence of root resorption from panoramic radiographs. Root resorption is a disease in which the cells of the tooth root are destroyed. Diagnosis of root resorption is made by clinical examination and a panoramic X-ray system, but computed tomography (CT) is necessary for accurate diagnosis. However, the current situation is that the diffusion rate of CT is low. In this study, we propose a method to extract features with fine-tuning using EfficientNet and learning using L2-softmax loss as the error function. We applied the proposed method to panoramic X-ray images of 59 real cases and obtained TPR: 71% and FPR: 43%. |
|
|
Classification of histological types of primary lung cancer from CT images using clinical information |
|
Naoya Honda, Tohru Kamiya(Kyushu Institute of Technology, Japan), Shoji Kido(Osaka University, Japan) |
| Accurate lung cancer identification is crucial, especially for fast-spreading small cell carcinomas that need early detection. Clinical data, like smoking history, is vital in computer aided diagnosis (CAD) systems alongside medical imaging. This study proposes a method to enhance accuracy by combining clinical information from medical records with images. Tumor images from CT scans are used, trained via deep learning, with a 5% accuracy boost observed when clinical data is added to 655 images of different cancer types. |
|
|
Recognition of Specific Parts of Plastic Bottles Using Improved DeepLab v3+ |
|
Daiki Ideta, Tohru Kamiya(Kyushu Institute of Technology, Japan) |
| In this paper, we focus on the manpower shortage in factories and conduct an experiment to try to automate the process. Among them, we focused on the sorting of plastic bottles at a waste disposal plant. In this paper, we focus on image processing technique based on deep learning. As basic research, we conducted an experiment to see how well the robot can identify a single plastic bottle in an image. We attempted to use semantic segmentation methods for detection, using DeepLabv3+ as the basic model, and improved it. scSE Block and fine-tuning were introduced to improve the accuracy significantly compared to previous studies. |
|
|
An Image Registration Technique for Brain MR Images Using Linear Transform by 3DCNN |
|
Seitaro Baba, Tohru Kamiya(Kyushu Institute of Technology, Japan) |
| In this paper, we focus on brain atrophy in Alzheimer's disease and propose a 3D image registration method of brain MR images for the purpose of calculating the atrophy rate by temporal subtraction technique. We improve the accuracy by using a linear 3D image registration CNN model and instance-specific Optimization. For the training of the model, we used pseudo-data obtained by arbitrarily linear transformation of the ADNI brain MR images. The experimental results show that the addition of the optimization process improves the NCC by more than 0.1. The loss function of the model is compared between Grid loss and MSE, and it is confirmed that MSE is effective for NCC. |
| |