基於斑馬魚運動軌跡之行為分類系統

Abstract

有關人類疾病的研究長久以來被人們所重視，由於斑馬魚基因與人類相近且可以被快速繁殖，斑馬魚在相關研究上扮演著重要的角色。斑馬魚經常被用來研究藥物與毒物對神經系統與行為的影響，斑馬魚已經被用來模擬過阿茲海默症、腦癌、癲癇、焦慮症以及肝臟疾病。然而目前研究斑馬魚的實驗經常需要仰賴肉眼或是昂貴商用軟體器材紀錄斑馬魚的運動軌跡，並做簡單的分析。因此進行斑馬魚實驗所需的時間、人力與金錢成本，使斑馬魚實驗變得複雜。本研究提出了一套斑馬魚行為模式分類系統Zebrafish Behavior Classification (ZBC) System，該系統能以軌跡資料對斑馬魚的行為進行分類。根據斑馬魚的啃咬、追逐、展示和正常等四種行為模式，我們提出了七種軌跡特徵的計算方法。這七種軌跡特徵包括軌跡相似度、總移動距離、平均幀間移動距離、最大和最小幀間移動距離、移動向量加總、最大和最小幀間移動距離向量，以及移動方向同方向的時間佔比。本系統會在計算七種軌跡特徵前先運用三種濾波方式清理軌跡資料，而七種軌跡特徵將用來訓練斑馬魚行為自動分類模型。

本研究比較了支持向量機 (Support Vector Machine, SVM)、隨機森林 (Random Forest, RF) 和極限梯度提升 (eXtreme Gradient Boosting, XGBoost) 等三種模型在斑馬魚行為自動分類模型上的準確度表現，同時我們也比較了均值濾波、中值濾波與卡爾曼濾波等三種資料清理方法運用在模型上的準確度表現。實驗結果顯示出ZBC在分類高度、中度和低度攻擊性的行為上能達到76%的準確度。另外，我們的研究發現不同資料清理方法和模型的搭配會影響分類的準確度，隨機森林在分類效果整體上有最高的準確度，而資料清理方法則須依照情況選擇。

Research in human diseases has long been a critical study, and zebrafish (Danio rerio) play an essential role in the field because of their genetic similarity to humans and can reproduce rapidly. Zebrafish are often used in experiments for the effects of drugs on behaviors, e.g., Alzheimer, brain cancer, epilepsy, anxiety, and liver disease. However, current experiments on zebrafish usually rely on the human eye or expensive commercial software to record the movement trajectory to analyze zebrafish behavior. As a result, the cost of zebrafish experiments in terms of time, labor, and money to make zebrafish experiments complicated to proceed.In this study, we proposed a Zebrafish Behavior Classification (ZBC) System to classify zebrafish behaviors. Seven trajectory features are founded based on four behavioral patterns of zebrafish, including biting, chasing, displaying, and normal behavior. The seven features are trajectory similarity, total moving distance, maximum and minimum moving distance per frame, average moving distance per frame, summation of moving vector, maximum and minimum angle of moving vector per frame, and same direction ratio for two trajectories. The proposed system employs three filters to clean up trajectory data before estimating seven trajectory features. The estimated seven features are used to train the Zebrafish Behavior Classification Model.This study compares the performance of accuracy on the Zebrafish Behavior Classification Model with three machine-learning methods - Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The performance of data processed by the three filters (i.e., Mean Filter, Median Filter, Kalman Filter) is also compared to show their results. The results show that the ZBC can achieve 76% accuracy in classifying zebrafish behaviors regarding the level of aggression behaviors - high, moderate, and low aggression. In addition, our study found that the combination of data-cleaning methods and machine-learning models would affect classification performance. In general, the random forest had the best classification results, and the choice of filters depended on the situation.