Prediction for Morphology and States of Stem Cell Colonies using a LSTM Network with Progressive Training Microscopy Images.

We present a new LSTM (P-LSTM: Progressive LSTM) network, aiming to predict morphology and states of cell colonies from time-lapse microscopy images. Apparent short-term changes occur in some types of time-lapse cell images. Therefore, long-term-memory dependent LSTM networks may not predict accurately. The P-LSTM network incorporates the images newly generated from cell imaging progressively into LSTM training to emphasize the LSTM short-term memory and thus improve the prediction accuracy. The new images are input into a buffer to be selected for batch training. For real-time processing, parallel computation is introduced to implement concurrent training and prediction on partitioned images.Two types of stem cell images were used to show effectiveness of the P-LSTM network. One is for tracking of ES cell colonies. The actual and predicted ES cell images possess similar colony areas and the same transitions of colony states (moving, merging or morphology changing), although the predicted colony mergers may delay in several time-steps. The other is for prediction of iPS cell reprogramming from the CD34+ human cord blood cells. The actual and predicted iPS cell images possess high similarity evaluated by the PSNR and SSIM similarity evaluation metrics, indicating the reprogramming iPS cell colony features and morphology can be accurately predicted.