Code for "Biophysical informatics reveals distinctive phenotypic signatures and functional diversity of single cell lineages"
Code is divided into two parts:
Cell instance segmentation - based on MASK-RCNN public code repo and adapted for cell images with custom NMS scripts.
Morphoscripts - morphological analysis, clustering, datavisualization, and plotting scripts that use segmentation output data.
Training a cell segmentation model first requires creating a dataset of images and segmentation annotations. This library accepts annotations in the standard pycocotools format, which can be prepared a number of ways (the authors have previously used Labelbox [https://labelbox.com/] and CVAT [https://cvat.org/] for this task). A path to the data and annotation json file can be provided to the python script data.py in line 203. If only a subset of the training images will be used during training, paths to specific images can be supplied by editing 'train_dataset' (line 210). Otherwise this can be left empty. Last, run data.py.
Running train.py will begin training. Running predict.py will run inference on a single image "in.jpg." In order to obtain analyzable segmentation results, run the python script model_predict_func.py. The output will be in the format imagename_instances.data file, which can be used directly in the downstream analysis.
Complete morphological analysis of a sequence of segmentation outputs corresponding to one cell line is most easily performed by initializing a linieage_timeless object, defined in the script lineage_timeless.py. The constructor takes in a filepath to a directory containing the outputs of the instance segmentation model (.data files) and calculates a range of properties at the single cell, cell network, and cell population over time levels. These are contained within the object (See the class definitions for lineage_timeless, network_object, and cell_object). As this computation is time consuming, we recommend saving the resulting files and reloading them for future use (through the save_object() and load_object() methods).
Cell network analysis is conducted using the network_object class, which is initialized by reading in a segmentation output. This precomputes a cell adjacency network as well as certain morphological features on a single cell and cell network basis. Additional morphological and topological features can be calculated using the methods defined.
Single cell morphology is likewise analyzed using cell network objects. If only cell morphological features are desired and not network topological features, a list of cell_objects can be created directly from the segmentation output file (loaded as 'instances') using:
for i in range(len(instances['pred_boxes'])):
cell_list.append(cell_object(instances['pred_boxes'][i],instances['pred_masks'][i],instances['scores'][i]))
For questions on how to run code, contact tjchan@seas.upenn.edu (510) 999 2031.