A deep learning approach to myofibroblast identification and segmentation

Introduction

Cellnet is a convolutional neural network trained to classify and segment 2 types of cells; activated myofibroblasts and fibroblasts.

The model input is a 3 channel flourescent image:

1. F-actin
2. alpha-smooth muscle actin
3. Nuclei / DAPI

The output is a segmented image with 3 classes:

1. Activated Myofibroblast
2. Nonactivated fibroblast
3. Background

Postprocessing provides the following for each image:

1. Number of cells
2. average activation of all cells
3. average area of all cells

Model Architecture

The base architecture of cellnet is a ResUnet with the encoder and decoder each consisting of 3 residual blocks. The model takes advantage of the F-actin channel of the input to segment a cell mask. This is then applied to the end of the decoder, effectivly eliminating the background class and reducing the number of predicted classes from 3 to 2.

Code

• application: for use applying a trained model to any data_set, can read images saved as numpy arrays or .nd2
• cell_utils: general utility functions for cellnet, including: model callbacks, metrics, and data generators
• create_dataset_aug: Used to generate training datasets including image augmentation
• post_utils: utility functions for application including cell mask and nuclei segmentation
• build_Resunet: functions used to build the model structure of cellnet
• traintest: Used to train all models, options to include attention, pvp, and squeeze and excite layers as well as a second output head