心筋細胞特異的マーカーであるNkx2.5発現細胞を蛍光蛋白質(GFP)で標識する遺伝子組み替えES細胞 (HES-Nkx2.5-GFP)を用いて、心筋細胞への分化誘導を行った。分化誘導は繰り返し行い、分化誘導過程において、時系列に位相差顕微鏡写真を撮影し、分化最終日に同時に蛍光画像を取得した。最初に分化誘導最終日の位相差顕微鏡画像から、心筋細胞の存在領域を免疫染色のように図示する仮想免疫染色システムを構築した。畳み込みニューラルネットワークを用いて、位相差顕微鏡画像中の心筋細胞を、Nkx2.5の蛍光画像を答えとすることで学習を行った。学習に使用していないテスト画像を用いて、画像を再構築し、人工着色を行った。結果、答えとなるNkx2.5の免疫染色と非常に類似した画像を作成することに成功し、心筋細胞の存在率に関しても高い相関(相関係数0.99)を持って予測可能であった。次に、心筋細胞への分化誘導初期段階の顕微鏡画像から、最終的に心筋細胞へと分化する画像であるか予測したいとか考えた。画像解析に最適である畳み込みニューラルと、時系列的なデータ解析に最適である再帰型ニューラルネットワークを組み合わせる独自のネットワークを構築し、学習を行った。学習の結果、分化誘導前の未分化ES細胞の状態から最終的な分化誘導を予測することが可能であった（AUC = 0.91）。未分化状態のES細胞において、既に最終的に心筋細胞へ分化するかどうか形態的な運命決定がされていると考えられた。分化誘導開始後 day1の時点で一度予測精度は低下するものの、day3の時点で再び予測精度は上昇し、時系列的な変化を加味することで高精度での予測が可能であった（AUC = 0.93）。分化誘導day3の時点で最終的に心筋細胞へと分化するコロニーが作成されたことが示唆される。同システムを活用することで、心筋細胞への分化誘導が成功するかどうか早期に予測することが可能となり、再生医療への応用が進むと考えられた。
We used a genetically modified ES cell (HES-Nkx2.5-GFP) that labels Nkx2.5-expressing cells, which are cardiomyocyte-specific markers, with fluorescent protein (GFP), and differentiated them into cardiomyocytes. Differentiation induction was repeated, and in the differentiation induction process, phase-contrast microscopic images were taken in chronological order, and fluorescence images were acquired at the same time on the final day of differentiation. First, a virtual immunostaining system was constructed in which the region where cardiomyocytes exist was illustrated like immunostaining from the phase-contrast microscopic image on the final day of cardiomyocytes differentiation. Using a convolutional neural network, cardiomyocytes in phase-contrast microscope images were learned by reference to fluorescent images of Nkx2.5 as an answer. Images were reconstructed and artificially colored using test images that were not used for learning. As a result, we succeeded in creating images very similar to the immunostaining of Nkx2.5, which is the answer, and it was predictable with a high correlation (correlation coefficient: 0.99) with respect to the abundance of cardiomyocytes. Next, we elucidated whether the cardiomyocytes differentiation at the final day can be predicted at the initial stage of differentiation from ESCs into cardiomyocytes, by using microscopic images. We constructed and trained a unique network that combines a convolutional neural network, which is optimal for image analysis, and a recurrent neural network, which is optimal for time-series data analysis. As a result of learning, it was possible to predict the final induction of differentiation from the state of undifferentiated ES cells before the induction of differentiation (AUC=0.91), which suggest that the morphological fate of undifferentiated ES cells has already been determined whether or not they will eventually differentiate into cardiomyocytes. Although the prediction accuracy decreased once on day 1 after the differentiation, the prediction accuracy increased again on day 3, and it was possible to make highly accurate predictions by using time-series analysis method (AUC = 0.93). It suggestes that colonies that finally differentiate into cardiomyocytes were created at the time of differentiation day3. By utilizing this system, regenerative medicine can advance, because we can predict whether cardiomyocyte differentiation would be successed at an early stage of induction into cardiomyocytes.