| LIANG TIANHENG | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 末次 志郎, | 小野 直亮, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
Title: Unsupervised Classification of Dilated/Hypertrophic Cardiomyopathy Using Pathological Images Abstract: The diagnosis of dilated/hypertrophic cardiomyopathy (DCM/HCM) using pathological images requires expertise in the field but the number of expert doctors is limited. Also it is a time-consuming task, highly depends on pathologists’ experience and knowledge. Several patch-based supervised learning approaches have been proposed to classify pathological images. However, given the patch diversity with respect to cellular contents, the assignment of a ground-truth label for each patch (i.e., patch-wise label), rather than a global (i.e. image-wise) label, is essential as it affects the training reliability. This project aims at leveraging unsupervised learning approaches to quantitatively assess the patch diversity with respect to disease classes. The proposed approach uses t-distributed stochastic neighbor embedding (t-SNE) and K-means clustering to assemble the patches into distinct clusters. In this study, 276 images from 96 cardiomyopathy patients were used. The images were split into patches with a matrix size of 256×256. Those patches including only the background were excluded, thus 7423 patches were used in the analysis. The patches were grouped into 300 clusters using t-SNE with a perplexity of 80. Four expert medical doctors were asked to classify the clusters into DCM, HCM or normal classes. The patch diversity was assessed by evaluating the agreement among the experts on the classification of each cluster. Results have shown that 36 clusters (12%) had agreement by all four experts, 130 clusters (43.3%) had agreement by three experts, and 64 (21.3%) clusters had agreement by two experts. These results indicate the intrinsic diversity in the patch classifications. | ||||||||
| 中西 直樹 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 向川 康博, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 | |
|
title: Decomposition of individual musculoskeletal of lower extremity from a radiograph using Deep Learnig
abstract: Decomposition of musculoskeletal structures of lower extremity from medical images is useful for quantitatively understanding the process of muscle atrophy caused by disease and the recovery process during the rehabilitation. In this study, we focus on decomposition of individual musculoskeletal structures from a radiograph that can be acquired with high spatial resolution and low radiation dose. Then, we treat the decomposition problem in a real radiograph as the image translation problem from a real radiograph to multiple digitally reconstructed radiographs generated using a CT image and the 3D mask of individual muscles which was obtained by a previously proposed automatic segmentation method. Since registration of real and synthetic radiographs is a challenging problem especially for the muscle structures, we introduce a image translation method based on unpaired training data set using CycleGAN. In this paper, simulation and real image experiments were conducted using CT images and real radiographs of 475 cases. language of the presentation: Japanese | ||||||||
| 本田 修平 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 向川 康博, | 荒牧 英治, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
title: Pre-training of the BERT model using a large-scale radiology report database for estimation of purpose of CT scanning
abstract: Our group has built a large-scale database consisting of CT images and radiology reports associated with each CT scan. In order to obtain knowledge related to anatomical variation from this database, we developed an automatic measurement system for musculoskeletal anatomical parameters using CNN and have analyzed changes in anatomical parameters such as pelvic tilt angle and muscle volume with age and gender. In this study, we attempted to estimate the objective of CT imaging (i.e., disease background) by using the radiological reading reports associated with each CT image to augment the medical knowledge obtained from the analysis results. The BERT model used in this study was pre-trained with a corpus of approximately 320,000 radiological reading reports and compared with the widely used Wikipedia corpus-based pre-training model. We created a criteria of classification of imaging purpose consisting of twelve classes and 299 reports were manually classified by an expert, which were used for fine-tuning of the pre-trained BERT model. The radiology report pre-trained BERT model improved the prediction accuracy by 18.1% over the conventional method using LSTM and by 7.7% over the Wikipedia corpus-based pre-training model. Finally, attention map was visualized to understand the rationale for the classification of the model, which demonstrated a reasonable agreement with expert's judgement for correctly classified reports. language of the presentation: Japanese | ||||||||
| CONG XI | M, 2回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 酒田 信親, | 磯山 直也 | ||
|
title: AR fire simulator based on environment recognition for fire protection
According to existing surveys, the main cause of the fire is improper handling of the fire source and combustible material. It is thought that the reason is that people are not aware of the danger of fire and lack of awareness of fire prevention. In this research, we aim to improve the user's awareness of fire prevention by showing user the state of a fire in the space which he or she lives every day by using a method of Augmented reality. Then, we introduced a general object recognition method to estimate the flammability recognized from the real world, Based on this, we will consider the addition of a function to simulate the diffusion process of fire. language of the presentation: Japanese | ||||||||