近年來,深度學習迅速地蓬勃發展,卷積神經網路(Convolution Neural Network, CNN)與遞歸神經網路(Recurrent Neural Network, RNN)分別在電腦視覺領域與時序列資料的處理展現各自的優點。遞歸神經網路(RNN)依賴先前時序列資料的特徵,並在許多領域的應用成功被證明是可行的方法,而RNN最常遭受到梯度消失或爆炸的問題。因此,之後提出的長短期記憶(Long Short-Term Memory, LSTM)與循環門控單元(Gated Recurrent Unit, GRU)的模型解決了RNN的問題。 本研究提出一套方法來改進時序列資料的預測。首先,我們提出累積式遞歸神經網路(Adaptive-Recurrent Neural Network, A-RNN)來提升更長遠的預測表現,將每回合新的預測結果遞歸至下一回合作為RNN的訓練資料,以進行下一個時序的預測。再者,基於A-RNN模型來訓練完整時序列的iPS細胞延時機率圖像且將它作為已訓練RNN模型,並應用於iPS延時顯微鏡圖像的細胞形成。已訓練RNN模型被分類成高成長型、中成長型及低成長型,而且分別是由各成長類型的完整時序列細胞訓練得來。本研究通過各成長模型的預測結果與實際圖像進行比較,選擇最適合的已訓練模型來做更準確的預測。最後,依照選擇的模型類型,而進一步推測新培養樣本的成長趨勢。
In recent years, deep learning has been popular rapidly. Convolution Neural Network (CNN) and Recurrent Neural Network (RNN) have each advantage in the field of computer vision and the processing of time-series data. RNN is dependent on the feature of previous data and has proved a successful method in many fields. However, RNN suffers from vanishing or exploding gradient problems. Therefore, Long Short-term Memory (LSTM) and Gated Recurrent Unit (GRU) solves the problem of RNN. This study proposes a method to improve time-series data predictions. First, we propose Adaptive-Recurrent Neural Network (A-RNN) to enhance long-term prediction performance and every new prediction is incorporated into every recursion of the RNN training data for the next prediction. Furthermore, based on the A-RNN model, which is trained from complete time-series data and used as the RNN trained model. It applies to time-lapse microscopy images during iPS formation. The trained RNN models are categorized as high-growth, mid-growth, and low-growth. There are trained from complete time-series cells of each growth type. This study compares the predicted results of each trained RNN model with the actual images, the most suitable trained model can be selected for accurate predictions. Finally, according to the type of the selected trained model, the newly iPS cell culture will be speculated the type of growing trend.