| 植田 秀樹 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: Inference of cleavage mechanisms of γ-secretase using machine learning
abstract: γ-Secretase is a membrane-embedded protease that cleaves amyloid precursor protein (APP) and generates amyloid beta protein (Aβ), which contributes to Alzheimer’s disease. Although γ-secretase is known to cleave APP successively, most likely in every three amino acids, the underlying mechanisms of the cleavage by γ-secretase are not well understood. In this research, we built a variety of machine learning models which predict the amount of a peptide cleaved out by γ-secretase. We selected a model by cleavage site predictions for the reported substrates. Based on the model, we inferred the number of pockets in the active site of γ-secretase, the physicochemical properties involved in cleavages by γ-secretase, and the conserved sequence of γ-secretase. language of the presentation: Japanese | |||
| 工藤 創大 | M, 2回目発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title:Analysis of medical deep learning models based on information bottleneck theory
abstract: Using data collected from signals from the Apple watch, a deep learning model was built to segregate signal patterns during sleep. Based on the information bottleneck theory, the deep learning model predicting COVID19 infection using CT images as well as the above model are analysed. This allows the relationship between prediction accuracy and information compression in each model to be measured. These results will be compared and discussed in terms of their practicality, including how much overlearning each problem has at present. language of the presentation: Japanese | |||
| 高松 拓実 | M, 2回目発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: Building an Embedding Model for Histopathological Images Using Deep Learning
abstract: In addition to the commonly used hematoxylin-eosin staining, various other staining methods exist for capturing histopathology images. The extracted tissue differs depending on the staining method, and the information obtained can be used to diagnose the type of tumor and evaluate its malignancy. However, since the images are cut from the same part of the body, the biological features are the same even if the staining is different. Therefore, the goal of this study is to construct a model that embeds different types of stained images into an integrated latent space, and to develop a method to extract biological tissue features independent of the image staining method. A Vector-Quantized Variational Auto-Encoder (VQ-VAE) was used as the learning model to build a model that embeds different types of stained images into an integrated latent space. In the future, I plan to verify whether it is possible to quantify and extract conditions such as tissue fibrosis and cell density using latent variables obtained from this model. language of the presentation: Japanese 発表題目:深層学習を用いた病理組織画像の埋め込みモデルの構築 発表概要:病理組織画像の撮影には一般的に使用されるヘマトキシリン・エオジン染色の他にも様々な染色方法が存在する。染色方法によって抽出される組織は異なり、そこから得られた情報により腫瘍の種類の診断、悪性度の評価などへの応用に利用できる。しかし、同一部分から切り出された画像であるため、染色が異なる場合でも生体的な特徴は同じであると言える。そこで本研究では種類の異なる染色画像を統合された潜在空間に埋め込むモデルを構築し、画像の染色方法によらない生体的な組織の特徴を抽出する手法を開発することを目標とする。学習モデルとしてVector-Quantized Variational Auto-Encoder(VQ-VAE)を用いて、種類の異なる染色画像を統合された潜在空間に埋め込むモデルを構築した。今後このモデルから得られる潜在変数を用いて組織の繊維化や細胞の密度といった状態を定量化し、抽出することが可能か検証を行う予定である。 | |||
| RUMMAN MAHFUJUL ISLAM | M, 2回目発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: Feature extraction and unsupervised clustering of cellular phenotypes in histopathological images of pancreatic cancer using information maximization
abstract: In recent years, computer-aided diagnosis based on deep learning concepts has become an attractive research topic in medical imaging. Most of these works utilized supervised learning methods that required prior pathological knowledge. However, it is necessary to extract potential features in images, based on unsupervised learning in order to obtain new pathological findings. For this reason, we implemented unsupervised cluster analysis based on maximization of mutual information to classify pathological images of pancreatic cancer into discrete categories. language of the presentation: English | |||
| LI ZHONGLUO | D, 中間発表 | 数理情報学 | 池田 和司, 向川 康博, 吉本 潤一郎, 福嶋 誠, 日永田 智絵 |
|
title: Attention-based Multi-Object Pose Estimation using Multi-task Network
abstract: Multi-Animal pose estimation is as difficult and challenging as in the case of human because it needs to figure out the number of animal and their positions. In addition, when the animals moving too rapidly or cross over by another mouse, most of deep learning models will fail to recognize it correctly. Attention-based architectures have been used for vision in recent years, for it leverages a self-attention matrix to aggregate features from arbitrary locations. On the other hand, we adapt multi-task framework to handle multi-object tracking task. Attention layer discriminate different individual and assign patches to the branch that belonged to recognition part. This system not only can analysis different video from the same experiment, but also easy to adpot to differnt experiment with similar object by transfer the feature recognition layer. language of the presentation: English | |||
| 浦 優輝 | M, 2回目発表 | 数理情報学 | 池田 和司, 佐藤 嘉伸, 吉本 潤一郎, 久保 孝富, 福嶋 誠, 日永田 智絵 |
|
title: Development of a Dog Emotion Estimation Model from Images Using Transfer Learning
abstract: Over the tens of thousands of years that dogs and humans have lived together, a strong emotional bond has developed between them. In recent years, dogs have been treated more like family members than pet animals, and research on dog emotions has been conducted actively. In this study, we focused on the estimation of dog emotion using dog images. If we can identify detailed emotions of dogs from dog images, it is expected to be useful for identifying the causes of problematic behaviors, detecting mental disorders, and improving the quality of life of dogs. Therefore, this study aims to develop a method for estimating detailed emotions from dog images. The training and evaluation of the model requires the creation of a dataset of dog images with detailed emotion labels. However, it is difficult to obtain a large amount of emotion-labeled data from experts. Therefore, we created a small-scale detailed emotion-labeled image dataset by having the model learn the features of dog images by solving a breed identification task in advance. We then tested the model's ability to estimate the detailed emotions of dogs by training it on a small dataset of detailed emotion-labeled images. The labeled dog emotion dataset was constructed by extracting images from multiple videos of a single dog, each labeled with 11 different emotions. We prepare two models: model A that has learned the features of general images, model B that has learned the features of dog images by transfer learning to the breed identification task from model A. We transfer them to the emotion estimation task. We compare and discuss four different models, one in which only the output layer is retrained and the other in which the entire model is retrained . language of the presentation: Japanese 発表題目: 転移学習を用いたイヌ画像によるイヌの情動推定モデルの開発 発表概要: イヌと人間は数万年にわたり共に生活する中で，両者の間には強い情動的な繋りが生じるようになった． 近年ではイヌは愛玩動物以上に家族のように扱われている． それに伴いイヌの情動に関する研究も盛んに行われている． 本研究ではイヌ画像を対象としたイヌの情動推定に着目した． イヌ画像からイヌの詳細な情動を識別できれば，問題行動の原因特定や精神疾患の発見，イヌのQOL改善などへの簡便な応用が期待される． そのため本研究ではイヌ画像から詳細な情動を推定する手法の開発を行う． モデルの学習・評価にはイヌの詳細な情動ラベル付き画像データセットの作成が必要となるが， 専門家による適切な情動ラベル付きデータを大量に確保することは困難である． そこで事前に犬種識別タスクを解かせることでイヌ画像の特徴をモデルに学習させた上で， 作成した小規模な詳細情動ラベル付き画像データセットにモデルを転移学習させることで，イヌの詳細な情動を推定するか検証を行った． 作成したイヌの情動ラベル付き画像データセットは11種類の情動がラベル付けされており，一匹のイヌを撮影した複数の動画から画像を切り出し構築した． 一般画像の特徴を学習済みのモデルと，そこから犬種識別タスクに転移学習しイヌ画像の特徴を学習させたモデルについて，それぞれ情動推定タスクへ転移学習させる． その際，出力層のみの再学習とモデル全体を再学習させた計4種類のモデルを比較し考察する． | |||
| 西村 虎太郎ジェームス | M, 2回目発表 | 数理情報学 | 池田 和司, 作村 諭一, 吉本 潤一郎, 久保 孝富, 福嶋 誠 |
|
title: SVM Boundary Correction by Local Density of Sample Distribution
abstract: In machine learning, one of the important factors is that the distribution of the data used for learning is balanced. However, in the real world, data sampled in experiments cannot always be expected to have an balanced distribution, and therefore, measures have been studied to deal with data with imbalanced distributions. One of the representative methods is the weight method, which assigns weights to the inverse of the frequencies of the majority and minority classes, respectively. However, the weight method ignores the local distribution of the samples because it assigns a constant weight to each sample. This may result in a lack of boundary correction. Therefore, we propose a classification method that focuses on the local distribution of samples. language of the presentation: Japanese 発表題目: サンプル分布の局所密度によるSVM境界補正 発表概要: 機械学習において，重要な要素の１つに学習に用いるデータの分布が均衡であるということが挙げられる．しかし，現実世界において実験等でサンプリングされたデータが常に均衡な分布をもつことは期待できない．そのため，不均衡な分布を持つデータに対する対策がこれまで研究されてきている．代表的な手法の1つに，多数派クラス，少数派クラスに対して各クラスの頻度の逆数を重みとして付与する重み法がある．しかし，重み法では各サンプルに対して一定の重みを付与するためサンプルの局所的な分布が無視される．これにより，境界が補正されない可能性がある．そこで本研究ではサンプルの局所的な分布に着目した分類手法を提案する． | |||