Supervised machine learning for automated classification of human Wharton's Jelly cells and mechanosensory hair cells

Abstract

Tissue engineering and gene therapy strategies offer new ways to repair permanent damage to mechanosensory hair cells (MHCs) by differentiating human Wharton's Jelly cells (HWJCs). Conventionally, these strategies require the classification of each cell as differentiated or undifferentiated. Automated classification tools, however, may serve as a novel method to rapidly classify these cells. In this paper, images from previous work, where HWJCs were differentiated into MHC-like cells, were examined. Various cell features were extracted from these images, and those which were pertinent to classification were identified. Different machine learning models were then developed, some using all extracted data and some using only certain features. To evaluate model performance, the area under the curve (AUC) of the receiver operating characteristic curve was primarily used. This paper found that limiting algorithms to certain features consistently improved performance. The top performing model, a voting classifier model consisting of two logistic regressions, a support vector machine, and a random forest classifier, obtained an AUC of 0.9638. Ultimately, this paper illustrates the viability of a novel machine learning pipeline to automate the classification of undifferentiated and differentiated cells. In the future, this research could aid in automated strategies that determine the viability of MHC-like cells after differentiation.

Fig 1. Process of feature extraction from images.

(a) Phase contrast scans taken from a prior experiment in which HWJCs were differentiated into MHC-like cellss. (b) An empirical gradient threshold (EGT) was applied to the phase contrast scan to separate the cells from the background. (c) A minimum cross entropy thresholding (MCET) method was applied to the EGT images to identify cells and then size, shape, and intensity features were automatically extracted.

Fig 2. Weights of all active features.

The absolute value of the weights of all active features, as determined by an L₁ regularized LR.

Fig 3. Variance of performance metrics.

The performance metrics for each model could vary significantly due to the randomness of the split of the data into the training and testing set. To account for this variation, the median scores for each performance metric across 1,000 iterations was used. The median was chosen due to the fact that the data were significantly skewed.

Fig 4. ROC curve of voting classifier and constituent models.

The ROC curve of the voting classifier model as compared to its constituent models.