Here we provide our code for the machine-learning based CHIR dose assessment and adjustment at stage I. We adopted the logistic regression model on hand-crafted features of 0-12h bright-field image streams.
The image stream data are saved as ./data/image/CD01-*/S*/T*.png, where CD01-* (CD01-1, CD01-2, CD01-3, CD01-4) is the batch name, S* (S1, S2, ..., S96) is the index of well, and T*.png (T1.png, T2.png, ..., T10.png) is the preprocessed bright-field image of the well at different time step of the image stream.
Then, run the following command
cd ./data
python compute_features.py
cd ..to obtain the image features (local_entropy_[0-9], cell_brightness_[0-9], fractal_dimension_[0-9], area_[0-9], circumference_[0-9], A_C_ratio_[0-9], optical_flow_[0-8]) for each image steam. These features will be saved as pandas DataFrame files (./data/CD01-*/CD01-*_features.pkl'); we also provide the *.csv file for an easy browse.
Finally, we convert the features to a 21-D feature vector for each image stream, obtain their ground-truth labels from the experimental results, and get the final dataset (also saved as pandas DataFrames, ./data/dataset.pkl). We further split the dataset into a training set (./data/dataset_train.pkl) and a test set (./data/dataset_test.pkl). Please refer to the jupyter notebook ./data/prepare_dataset.ipynb in python for more detail.
Note:
Since the total size of the image stream data is extremely large (~52GB), we only provide the computed features for each image (./data/CD01-*/CD01-*_features.pkl). If you want to reproduce our feature extraction, please contact the corresponding author for the image stream data.
We first visualize the feature space using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) (Fig. 4d, Supplementary Fig. S10c-e). We provide the code for generating the LDA/PCA plots in the jupyter notebook ./feature_visualization.ipynb.
We next perform machine learning to classify each well as "low", "optimal" or "high" under CHIR duration = 24h, 36h, and 48h. We evaluate the classification performance by accuracy, precision, recall, F1-score, and AUC (Fig. 4e, Supplementary Fig. S10f). We also derive the feature importance weights by ANOVA (Supplementary Fig. S10b). To intervene in the differentiation condition in time, the classifier can also provide a "Deviation Score" for each CHIR dose (Fig. 4f). The predicted Deviation Scores can help select the optimal CHIR duration and improve the differentiation efficiency (Fig. 4h). The relevant codes are provided in the jupyter notebook ./machine_learning.ipynb.
To test the model’s generalization ability to new batches, we conduct a cross-batch validation under a CHIR duration of 24h (Fig. 4g, Supplementary Fig. S10g, h). The code for cross-batch validation is provided in the jupyter notebook ./cross_batch_validation.ipynb.