摘要 近年來,人類誘導性多能幹細胞 (Human Induced Pluripotent Stem Cells, hiPSCs) 的成功培養使得再生醫學有著突破性的發展,但是由於其培養過程面臨細胞誘導效率不佳等問題,導致 hiPSCs 的培養需耗費大量的時間與成本。 本研究利用深度學習方法,結合 UNet 對延時顯微鏡影像使用卷積神經網路 (Convolutional Neural Network, CNN),針對 hiPSCs 的培養過程進行分析,讓我們可於較早期時序預測 hiPSCs 的形成與否,並基於 CNN 部分時序列的 hiPSCs 機率影像結合遞歸神經網路 (Recurrent Neural Network, RNN),針對 hiPSCs 的生長進行預測,讓我們可以透過預測的分析來了解 hiPSCs 的生長趨勢。 透過針對多組延時顯微鏡影像的實驗結果顯示,本研究這兩套對於 hiPSCs 的預測機制有助於在早期評估 hiPSCs 的培養狀態,並根據對其趨勢的分析來預測 hiPSCs 的生長狀況,將可用於進一步決定是否要繼續培養 hiPSCs。
Abstract In recent years, the successful cultivation of human induced pluripotent stem cells (hiPSCs) has led to a breakthrough in regenerative medicine, but the cultivation of hiPSCs is time-consuming and costly due to the poor cell induction efficiency of the culture process. In this study, we use deep learning methods on time-lapse microscopy images by combining UNet and Convolutional Neural Network (CNN) to analyze the culture process of hiPSCs which allows us to predict the formation of hiPSCs at an earlier time sequence and by combining CNN probability maps and Recurrent Neural Network (RNN) to predict the growth of hiPSCs, allowing us to understand the growth trend of hiPSCs through predictive analysis. The experimental results of this study showed that the two predictive mechanisms of hiPSCs can help to assess the culture status of hiPSCs at an early stage and predict the growth status of hiPSCs based on the analysis of their trend, which will be used to further decide whether to continue culturing the cells or not.