| 政木 勇人 | D, 中間発表 | 生体医用画像 | 佐藤 嘉伸, | 金谷 重彦, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
title Automatic musculoskeletal segmentation from whole-body CT images integrated with individual labeled CT image databases
abstract The segmentation of the whole-body musculoskeletal system from CT image is a difficult task due to the high cost of creating annotations, and variability in the field-of-view (FOV) covered at large CT image databases. In the past previous studies, several whole-body skeleton segmentation methods have been proposed, but as far as we know, detailed whole-body musculoskeletal segmentation including annotation of individual muscles has not been performed. In this study, in order to reduce the annotation cost, we aim to develop an automatic musculoskeletal segmentation method from whole-body CT images by effectively integrating the various FOV training databases of separately and individually existing anatomical structures. In this study, we conducted preliminary experiments for validating a method for integrating outputs of multiple models trained on various training databases. language of the presentation Japanese | |||||||
| 伊東 尚輝 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 金谷 重彦, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
title Musculoskeletal Segmentation of the Lower Limbs to Reduce the Effect of Posture in CT Imaging
abstract Patient-specific musculoskeletal analysis can assist in the orthopedic diagnostic and surgical planning procedures. In our previous study, we achieved high accuracy muscle segmentation at the region around the hip joint. It is an important issue to extend this method to the entire lower limb. However, the posture when taking CT images varies depending on the medical institution and individual patient conditions. Therefore, there are variations in the postures in the dataset that lead to degraded performance when segmented using a single deep learning model covering the whole region. To solve this problem, we propose a method that segments a CT volume through dividing it into multiple regions of interest (ROI) and performs learning and inference in an ROI-wise manner. In this study, as a preliminary experiment, we validated the proposed method on CT images of the muscles around the hip joint. language of the presentation Japanese | |||||||
| 成田 剛志 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 加藤 博一, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
title Improving the Accuracy of Musculoskeletal Segmentation of CT Images from Multiple Institutions with Different Imaging Conditions
abstract The automatic segmentation of lower limb musculoskeletal structures from CT images is important in diagnostic support systems for orthopedic surgery and prognosis. Although automatic segmentation methods based on deep learning have been proposed for the automatic segmentation of lower limb musculoskeletal structures, their segmentation performance is degraded by differences in imaging modalities and imaging conditions for images acquired from multiple institutions. In order to improve the segmentation accuracy by reducing the effect of degraded image quality, we propose two methods: 1)Using the estimation uncertainty produced by multiple predictions via Test-Time-Augmentation (TTA), and 2) Using the domain translation of input images by Pix2Pix network. In the experiments, we performed quantitative and qualitative evaluations of the segmentation accuracy using test data having different imaging conditions than those of the training data. language of the presentation Japanese | |||||||
| ZHANG BIN | 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 Evaluation of Bayesian Active Learning for Segmentation of Liver and Spleen in Large Scale Abdominal MR Data Set
abstract Manual annotation in image segmentation is time-consuming and expensive. In order to obtain large number of annotated data set efficiently, Bayesian active learning has been proposed. The key component in the iteration in Bayesian active learning is the selection of query slices (or voxels) which maximize the performance of the model trained in the next iteration. We need to take account for (1) uncertainty estimated from the model trained in the previous iteration, i.e., the distance from the existing training data set, and (2) similarity among the query images. The large batch acquisition with diverse images far from the existing data set enables higher efficiency in active learning. In this study, we investigated the performance and efficiency of several Bayesian active learning approaches specifically for segmentation of liver and spleen in a realistic simulation study using 251 fully annotated abdominal MR data set. language of the presentation English | |||||||