- 學位論文
以AI技術應用於細胞行為之分析
Cellular Behavior Analysis through Artificial Intelligence Technology
摘要
幹細胞品質的好壞是決定再生醫學治療效果的關鍵因素，幹細胞在連續的繼代培養下，衰老的細胞在基因表現上逐漸不穩定且具有細胞轉變的風險，因此，幹細胞療法不宜選用衰老的細胞進行治療，然而，目前沒有完善的特異性標記可以評估衰老的細胞。根據先前的研究結果顯示，細胞的形態與遷移行為可以反映傳代效應的幹性變化，於是，基於細胞的形態特徵及動態遷移行為指標，本研究提出一套幹細胞培養品質的即時鑑別輔助系統，利用機器學習、深度學習技術進行幹細胞種類、培養代數的識別預測，以達到初步有效的體外培養品質監測。結果顯示，應用機器學習對不同代數下的細胞種類識別中，以成熟代數具有最好的識別成效，準確率為96.3%；而在不同種類細胞的培養代數識別，以hPDMCs具有最好的識別成效，準確率為81.7%。相較之下，應用深度學習對細胞種類的識別中，成熟、老年代數的種類識別準確率皆達95.0%以上；而在培養代數識別中，以hPDMCs具有最好的識別成效，準確率為92.6%。因此，本研究證實運用機器學習及深度學習所建立的細胞種類或培養代數的識別模型，對細胞品質監測具有可行性。未來將可增加細胞的種類或培養代數，並對訓練模型的架構進一步調整，期望能提高準確率並藉此瞭解各個辨別指標與傳代間的關聯性。
並列摘要
The quality of stem cells is the pivotal factor in determining the therapeutic effect of regenerative medicine. In serial sub-cultivation, senescent cells are inadequate for stem cell therapy. However, there is currently no real-time tool to assess stem cells getting senescent in situ. According to previous study, the morphology and migration behavior of cells could reflect the stemness changes of the passaging effect. Therefore, based on the morphological characteristics and migratory behavior indicators of the cells, this study proposed a computer-aided system to monitor stem cell quality during in-vitro culture by applying the machine learning and deep learning technologies. Applying machine learning to identify cell type at different passages, the highest accuracy was 96.3% at matured passage. In addition, applying machine learning to classify cell passage of each cell type, the results showed up to 81.7% accuracy for hPDMCs. By contrast, applying deep learning to identify cell type at different passages, the accuracy could reach more than 95% at matured and degenerated passages. In addition, applying deep learning to classify cell passage of each cell type, the highest accuracy was 92.6% for hPDMCs. Therefore, this study proved the feasibility of the model established by machine learning and deep learning to identify cell type or passage in monitoring the quality of stem cell during in vitro culture. In the future, the architecture of the training model will be further modified by adding more cell types or passages, hoping to improve the accuracy and to understand the relation between each classification feature and cell passage.