Deep Learning Solution to TIMING (Time-lapse Microscopy in Nanowell Grids) Project created and maintained by Single-Cell Lab, FARSight Group, STIM Laboratory and HULA Lab at University of Houston, Houston, Texas.
Deep learning algorithms show top performance on image classification and object detection tasks over general-purpose image domains (IMAGENET, MSCOCO). This repository explores the performance of state-of-art image classifier and object detector on a different domain consisting of time-lapse microscopy images collected in TIMING project. Extended from the initial exploration, we created annotated TIMING datasets efficiently using man-in-the-loop fashion and our customized annotation tool, to support the fine-tuning of the deep learning models. And an end-to-end TIMING pipeline is developed to detect, track and quantify cells with robustness and flexibility.
-
High-quality annotated TIMING Data. Leveraging unsupervised algorithm and man-in-the-loop fashion, we generated more than 5,000 nanowell images with cell bounding box annotations, 180 sequences of time-lapse nanowell frames with bounding box as well as track ID annotations and 72,000 cropped cell patches with apoptosis status (positive/negative) annotations. Annotated data can be found in the Deep TIMING Supplementary Materials folder.
-
Cell detection leveraging phase-contrast channel and Faster R-CNN algorithm. Trained state-of-art cell detector using Faster R-CNN model. Tested cell detection performance using different combination of input channels.
- Label-free apoptosis classification using CNN and LSTM Models Implemented state-of-art convolutional neural net classifier. Used reccurent model (LSTM) to improve sequence classification performance considering temporal dependency.
- Customized visualization tool from TIMING2-board
- 64-bit computer with at least 2GHz processor running Windows, Linux or Mac
- CUDA-enabled GPU, memory >= 8 GB recommended
- Hard drive storage >= 2TB, solid-state hard drive strongly recommended
(1) Download this repository and put the folder say /HOME_DIRECTORY/DEEP-TIMING/
(2) Download auxiliary modules and test data, and copy the folders to /HOME_DIRECTORY/DEEP-TIMING/ and to /HOME_DIRECTORY/DEEP-TIMING/DATA/raw/. (create folder DATA and subdirectory /raw/ and /results respectively.)
(3) Download and install Anaconda
(4) Create the environments for TIMING2-pipeline and TIMING2-board, open Anaconda Prompt, change to DEEP-TIMING home directory /HOME_DIRECTORY/DEEP-TIMING/, and type python setup_env.py
(5) Set up DT-pipeline, in the prompt, type activate DT-pipeline, and then type python setup_DT_pipeline.py
(6) Open another Anaconda Prompt, change to DEEP-TIMING home directory, type activate DT-board, and then type python setup_DT_board.py (independent from step 5)
(7) Have a cup of coffee, will be ready in several minutes.
1.DT-Pipeline wil do cell detection, tracking and feature calculation with phase contrast channels; follow the steps in Deep-TIMING-Pipeline-Demo.ipynb;
2.The evaluation data, scripts are included in the Supplementary Materials. And the evaluation steps can be fould here
Email: hlu9@uh.edu
This code is free for non-commerical use only