Automated phenotypic analysis and classification of drug-treated cardiomyocytes via synergized time-lapse holographic...

Predicting cardiovascular risk is critical for the therapy and control of cardiovascular illnesses. This work studies screening the toxicity of three drugs, (E-4031, isoprenaline, and sertindole) with various concentrations using tracking of the single cardiac cell's contractile motion to explore the drug concentration impact on single cell contractile dynamics and automated classification based on cells motion behavior using deep transfer learning and different machine learning based methods.

We developed an integrated platform for automated dynamic analysis and classification of human-induced pluripotent stem cell-derived cardiomyocytes (CM) at the single-cell level that uses a combination of holographic image-based tracking and deep learning. For in-depth investigation and automated classification of CMs, first, we accurately extracted a single CM's dry mass using time-lapse holographic imaging and a deep fully convolutional network. Afterward, we applied the Farneback optical flow method to track the cell's contractile motion frame by frame through the extracted single CMs with single-pixel displacement detection. Following this, a computational algorithm was developed to further characterize the single CM's functional behaviors, and several beating activities-related parameters were calculated. The average result of the population for every measured parameter was compared to the control condition using an unpaired student t-test. Finally, we proposed a fine-tuned deep transfer learning-based model for automated cell classification based on the compound's mode of action from the single-cell motion waveform and compared the result to feature-based single-cell classification using different machine learning approaches. We also provided reliable synchronization analysis of drug-treated CMs owing to cellular analysis at the single-cell level.

Cardiomyocytes responded to isoprenaline by speeding up their action potentials, increasing the frequency of their beats. The action potential speed was reduced and the resting time was lengthened in the presence of E-4031 and sertindole lowering the beat frequency. Performance analysis of the deep transfer learning-based model was carried out using several well-known evaluation metrics. A performance analysis shows the proposed model achieved 98.8 % classification accuracy.

Our implementation allowed precise single drug-treated CM motion tracking and in-depth investigation and automated classification of CMs via synergized time-lapse holographic imaging and deep learning.