コロキアムB発表

日時: 6月27日(月)3限(13:30-15:00)


会場: L2

司会: 北野 和哉
CHEN YINGDONG D, 中間発表 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: CNN-based automated segmentation of deep vessels in lower limbs from non-contrast-enhanced CT images
abstract: Deep vessels in lower limbs pass close to the pelvis in many patients with hip diseases, and incidences of vascular injury during surgery in the hip have been reported. Careful preoperative planning taking into account the vessel location is necessary to avoid vascular injury. Manual definition of the vessels in non-contrast-enhanced CT images, which are mainly used for hip surgery planning, is tedious due to the low contrast between vessels and neighboring structures. Therefore, the automated segmentation of deep vessels in these images would be helpful. In this study, we applied CNN-based U-Net to the segmentation of deep vessels from non-contrast CT images. Our approach of the segmentation consists of three steps: 1) Training and validation using 20 manually annotated clinical CTs with and without neighboring muscle information 2) Quantitative evaluation of vessels manually defined and semi-automatically defined from contrast-enhanced images 3) assessment of the distance between the vessels and the acetabulum rim (surgical target) in the hip region. Our experiments show the effectiveness and feasibility of CNNs in deep vessels segmentation and distance assessment from non-contrast enhanced CTs.
language of the presentation: English
 
LI GANPING 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: Construction of large-scale MR image dataset based on Bayesian Active learning
abstract: Image segmentation is a fundamental problem as it extracts the organ/tissue of medical interest for a series of medical researches, including the aging of the musculoskeletal system. Current advances in supervised learning have achieved promising results on many medical image segmentation tasks. However, it requires large numbers of annotated data to utilize these methods on musculoskeletal segmentation due to the large variation among medical images (variations in anatomy, inter-structure difference, noise). In contrast, manual annotations for medical image needs much effort and cost. In this study, we estimate the feasibility of Bayesian U-net with Monte Carlo dropout and uncertainty metric in active learning with 110-case partially annotated MR dataset and built a large-scale lower limb MR image dataset of 136 cases with muscle-wise refinement based on 3 reivsed cross-modality prediction in our previous work. The result showed that Bayesian active learning achieves in acquiring samples of high variety in limited batches and significantly reduces annotation effort. A large-scale longitudinal dataset of 3-year interval will be constructed with an improved sample selection policy and refinement method in future for a further clinical study.
language of the presentation: English
 
GU YI 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, 久保 孝富
title: Musculoskeletal Decomposition and High-fidelity Body Composition From a Plain X-ray Image
abstract: Musculoskeletal Health is becoming inreasingly important, espevially in establishing a sustainable super-aging society. Osteoprosis and sarcopenia are prevalent musculoskeletal diseases causing bone fracture, which leads to decline in activity of daily living. To frequently monitor musculoskeletal health, low-cost, low-radiation dose, and ubiquitously available but accurate musculoskeletal diagnosis method is highly anticipated. This study presents methods of high-fidelity body composition used to diagnose musculoskeletal disease from a plain x-ray image through decomposition. In bone mineral density (BMD) experiments, the proposed method achieved Pearson correlation coefficient (PCC) of 0.882 and 0.887 on a 200-cases datset and 748-cases multi-center dataset, respectively. In muscle density experiments, the proposed method achieved intra-class correlation coefficient (ICC) of 0.706, 0.713, 0.751, and 0.745 on gluteus medius, gluteus minimus, iliacus, and psoas major muscles, respectively. Future works include improving estimation accuracy and performing large-scalre clinical validation.
language of the presentation: English
 
小出 ゆり M, 2回目発表 計算システムズ生物学 金谷 重彦, 宮尾 知幸, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘
title: Classification of metabolites by metabolic pathways concerning to terpenoids, phenylpropanoids and polyketide compounds based on machine learning
abstract: Terpenoids, phenylpropanoids, and polyketides are the majority of the secondary metabolites containing carbon, hydrogen, and oxygen. In this work, 19,769 metabolites accumulated in KNApSAcK Core DB were classified into 71 subgroups comprising three major groups (terpenoids, phenylpropanoids, and polyketides) according to scientific literatures. We represented the metabolites as molecular fingerprint including chemical properties, and used those descriptors for classification by random forest model. We found that both training and test metabolites were well classified into the subgroups, with 94.06 %, and 94.23 % accuracy, respectively. Though classification of metabolites based on metabolic pathways is very time-consuming works, machine learnings with molecular fingerprint made it possible to attain the classification. This work will lead a light for systematical and evolutional understanding of diverged secondary metabolites based on secondary metabolic pathways. Data science is an interdisciplinary and applied field that uses techniques and theories drawn from statistics, mathematics, computer science, and information science. Combining these resources data science enables extracting meaningful and practical insights for secondary metabolites.
language of the presentation: Japanese