Poster M1 - AI-Based Recognition of Single Beta Cells in Heterogeneous iPSC-Derived Cell Populations Using Brightfield Images

Introduction: Beta cells, which produce insulin, are found in pancreatic islets, which are scattered cell clusters. Beta cell dysfunction or depletion can cause diabetes mellitus, which impairs blood glucose regulation. Diabetes is one of the fastest-growing global diseases, affecting 643 million by 2030 and 783 million by 2045. Insulin-dependent diabetic patients may benefit from islet transplantation. This method faces significant obstacles despite its potential. The lack of donors and immunosuppressive issues make it difficult to deploy, stressing the need for novel diabetic treatment methods. These issues can be addressed by creating islet organoids. Human induced pluripotent stem cells can be used to create organoids, three-dimensional cell aggregates that mimic organ form and function. One restriction of organoid differentiation is its stochastic nature, resulting in a diverse cell population after the ~month-long process. Organoids must undergo invasive, time-consuming phenotypic testing, which is not possible for real-time characterization. Here, we present an innovative and non-invasive technology that helps us to conduct an accurate selection of desirable cells after the final stage of differentiation among the heterogenous cell population to produce high quality islet organoids. We hypothesize that brightfield images of  cells derived from hiPSCs contain adequate information, which can be extracted using convolutional neural network (CNN), to establish the unique morphometric signatures associated with differentiation. CNN is a type of deep learning (DL), which is an algorithm in machine learning, used in computer vision. We developed a very accurate CNN model specifically designed for beta cells derived from hiPSCs.

Materials and

Methods: Image Acquisition:
ImageStream Mark II, INSPIRE 744 using 60X magnification, was utilized for the image acquisition. The gate for positive and negative cells was positioned using the fluorescent images based on the expression of the markers.

Criteria for Cell Classification:
To establish the ground truth labels, we used the fluorescent images of the transfected and WT cells to generate two distinct classes (negative, and positive) for the binary classification task. We utilized the expression of the m-Cherry-INS reporter in the transfected cells and NKX-6.1 and C-peptide in the WT cells to select the bright-field images for the positive class.

Dataset Preprocessing:
The transfected (2,274 images) and wild-type (99,886 images) datasets were separated into three subsets: training (80%), validation (10%), and test (10%). Transfected and WT cell images were scaled to 224 by 224 pixels utilizing zero-padding around images and row and column replication at the farthest edge respectively before feeding into the neural network.

Threshold of CNN Model:
The neural network calculated the input data probability class using the softmax function in the output layer. The thresholds were chosen by computing the median value and using histogram of the probability of the predicted outcome. We used the ResNet architecture and optimized the hyperparameters using the validation dataset.

Performance of CNN models:
The evaluation of our models was conducted using accuracy, recall, precision, F-measure, Mathews correlation coefficient (MCC), and the area under the receiver operating characteristic curve (AUCROC). Grad-CAM was used to visualize important regions of the single cell images.

Results, Conclusions, and Discussions: Results
For the proof of concept, we built a CNN model using cell images that contained a reporter gene for insulin. Building the CNN model required a 32-batch size, 0.00001 learning rate, and 0.9 momentum. The model performed well after training, with 0.97 accuracy, 0.97 F1 score, and 0.99 ROC-AUC. Then, we developed a CNN model using WT cell images and we adjusted the hyperparameters to 32-batch size, 0.00001 learning rate, and 0.9 momentum. The properly trained model displayed loss values around 0.2 and no discernible difference between the loss values in the training and validation data (Figure 1). Our model did well after training, scoring 0.92 accuracy, 0.923 F1 score, and 0.971 ROC-AUC (Figure 2,3). We chose two thresholds (0.69 and 0.96), based on the statistical description and SoftMax probability histogram. Our model obtained 0.92 accuracy, and 0.92 F1 score at 0.69. At 0.96 threshold, the model had 0.74 accuracy, and 0.65 F1 score.
Discussion
To verify our hypothesis, we built a model using m-Cherry-INS cell images that classified well the bright-field images, indicating the possibility to capture morphological signature of the differentiated cells. Then, we developed a model capable of identified the β cells derived from WT iPSCs. The majority of the images were accurately identified as either positive or negative by our model, and it may be fine-tuned to minimize false positive predictions. At a threshold of 0.69, the model achieved a 2% reduction in false positives compared to a threshold of 0.5. At a threshold of 0.96, the model achieved a 5% reduction in false positives, but this came at the expense of the model's overall accuracy. The GRAD-CAM heatmap demonstrates that the cell image insight pixels were the most important ones for the model's choice.
Conclusions
In this study, we demonstrated that enough information is provided by the brightfield images of β cells derived from hiPSCs to distinguish the unique morphometric traits associated to differentiation. The approach works with any kind of cell derived from iPSCs and is suitable for large-scale manufacture. Our AI model is very adaptable and capable of meeting clinical requirements.

Acknowledgements (Optional):