Skip to content
This repository was archived by the owner on Jan 25, 2018. It is now read-only.

Home

Jump to bottom
Virginie edited this page Jan 16, 2015 · 7 revisions

Welcome to the CPCharm wiki!

Repository content

Modules folder:
CP-CHARM-specific modules that are not yet included in latest CellProfiler release. The use of these modules outside of CP-CHARM pipelines is not recommended before their inclusion in CellProfiler's official release.

Pipelines folder:
Ready-to-use CellProfiler pipelines for extraction of the CHARM-like feature vector.

Classifier folder:
Python scripts for training and testing a classifier, as well as for further label-free classification.

Example folder:
Application example of CP-CHARM. The folder contains:

  • Input_Images, a folder containing images from a 2-class problem (Negative and Positive)
  • DefaultOUT_CHARM-like_training_data.csv and DefaultOUT_CHARM-like_training_labels.csv, the feature vectors and corresponding labels for all images extracted using the [TrainTestMode]CHARM-like pipeline
  • pcalda_classifier.pk, a pca-lda classifier trained and tested using 10-fold cross validation with the measurements from DefaultOUT_CHARM-like_training_data.csv
  • results_summary.txt, a report of pcalda_classifier.pk training and testing phase
  • wnd_classifier.pk, a WND classifier trained and tested using WND-CHARM's custom validation method with the measurements from DefaultOUT_CHARM-like_training_data.csv
  • predicted_labels.csv, the output of blind classification of Input_Images using the previously trained and tested pcalda_classifier.pk classifier

Setting up CP-CHARM-like

  1. Download compiled version of CellProfiler
  • Pipelines have been designed for two different CellProfile version, which can be downloaded at http://www.cellprofiler.org/previousReleases.shtml
    • [TrainTestMode]CHARM-like.cp and [ClassifyMode]CHARM-like.cp are designed for CellProfiler release version 2.0.11710
    • [TrainTestMode]CHARM-like.cppipe and [ClassifyMode]CHARM-like.cppipe are designed for CellProfiler release version 2.1.0

We advise to use the two pipelines designed with CellProfiler version 2.1.0, namely [TrainTestMode]CHARM-like.cppipe and [ClassifyMode]CHARM-like.cppipe.

We are currently updating the pipelines to CellProfiler's latest release, but in the meantime we recommend only using them with their native CellProfiler version.

  1. Install modules specific to CP-CHARM
  • Download the Modules folder with all its content
  • In CellProfiler's Preference menu, set up the plugin directory to point to your downloaded Modules folder
  • Note that you must restart CellProfiler after modifying the plugin directory path

Running CP-CHARM-like

  1. Extracting features in CellProfiler
  • Two pipelines (per version) are available, located in the Pipelines folder:

    • [TrainTestMode]CHARM-like gathers metadata on filename (image ID), folder name (class), and eventual batch information (holdout), and outputs two .csv files: DefaultOUT_CHARM-like_training_data.csv and DefaultOUT_CHARM-like_training_labels.csv

      • DefaultOUT_CHARM-like_training_data.csv contains 954 columns (953 feature measurements plus a unique image identifier) and I+1 rows, where I is the total number of analysed images. The first row is composed of identifiers for each feature measurement (feature name) and the Metadata_Key field (indicating the column containing unique image identifiers).
      • DefaultOUT_CHARM-like_training_labels.csv contain 2 or 3 columns depending on the experiment and I+1 rows, where I is the total number of analysed images. The first row is always composed of the Metadata_Key (indicating the column containing unique image identifiers) and the Metadata_Class (indicating the column containing class labels) fields, and can also have a Metadata_HoldOut (indicating images from the same batch group) field.

    • [ClassifyMode]CHARM-like gathers metadata on filename (image ID) only and outputs one .csv file: DefaultOUT_CHARM-like_data.csv

Note that the content of Metadata_Key, Metadata_Class and Metadata_HoldOut has to be adapted to the file naming system of the studied dataset by creating appropriate regular expressions in LoadImages (first module of the pipeline). Metadata_Key must contain strings that uniquely identify each image. Metadata_Class must contain the same string for all images of the same class, and be different for images of different classes. Finally, Metadata_HoldOut must contain similar identifiers for images coming from the same batch, which differ from identifiers of images from other batches.

In the training and testing phase, the fields Metadata_Key and Metadata_Class are mandatory to uniquely identify each image in the training set as well as its corresponding class. The field Metadata_HoldOut is not required if the [HOLD OUT SAMPLES]option is set to false (0) during classification. This field allows to ensure safe splitting of training and testing set to minimize batch effect during validation. In blind classification phase, only the Metadata_Key is required.

  • For either pipeline, either open CP, load the CHARM-like pipeline you want and run analysis, or run CP headless (from CellProfiler directory) using the following command:

      python CellProfiler.py -p ../CHARM-like.cp -c -r -o ../ -i ../Input_Images --plugins-directory=../Modules/
  1. Training and testing using traintest.py script, located in the Classifier folder

  • Severals options are available using the following command line (from the Classifier folder):

      python traintest.py [DATA_FILE (.csv)] [LABELS_FILE (.csv)] [OUTPUT_PATH] [NB_RUNS] [HOLD OUT SAMPLES (1:yes, 0:no)] [DISPLAY CONFUSION MATRIX (1:yes, 0:no)] [CLASSIFICATION_METHOD ("lda", "wnd")] [VALIDATION_METHOD ("save25", "kfold")] [NB_FOLDS (if VALIDATION_METHOD = "kfold")]

  • To run CP-CHARM-like with the default output from the [TrainTestMode]CHARM-like.cppipe CP pipeline, use:

      python traintest.py ../DefaultOUT_CHARM-like_data.csv ../DefaultOUT_CHARM-like_labels.csv [OUTPUT_PATH] [NB_RUNS] [DISPLAY CONFUSION MATRIX (1:yes, 0:no)] "lda" "kfold" [NB_FOLDS]

Where [NB_RUNS] is the number of times you would like the training/validation to repeated and [NB_FOLDS] is the number of folds in k-fold cross-validation. [OUTPUT_PATH] is the path to the directory here you would like the program to output the results (results_summary.txt) and the classifier (wnd_classifier.pk or pcalda_classifier.pk)

  • If one wants to save the confusion matrices in addition to the default results_summary.txt (which contains classification accuracies and the list of parameters used to do the experiment), [DISPLAY CONFUSION MATRIX] should be set to 1 and everything should be piped in a text file as in the following example:

      python traintest.py ../DefaultOUT_CHARM-like_data.csv ../DefaultOUT_CHARM-like_labels.csv [OUTPUT_PATH] [NB_RUNS] 1 "lda" "kfold" [NB_FOLDS] > ../confusion_matrices.txt

  • Here's an example of a command that will run 10 rounds of 10-fold cross-validation using PCA-LDA and save the confusion matrices:

      python traintest.py ../DefaultOUT_CHARM-like_data.csv ../DefaultOUT_CHARM-like_labels.csv ../ 10 1 "lda" "kfold" 10 > ../confusion_matrices.txt

  • Here's one that will run 100 rounds of "leave 25% out" validation using WND-CHARM without saving confusion matrices:

      python traintest.py ../DefaultOUT_CHARM-like_data.csv ../DefaultOUT_CHARM-like_labels.csv ../ 100 0 "wnd" "save25"
  1. Classifying using classify.py, located in the Classifier folder

  • Use the following command:

      python classify.py [CLASSIFIER_FILE (.pk)] [DATA_FILE (.csv)] [OUTPUT_PATH]

Where [CLASSIFIER_FILE (.pk)] is the classifier outputted by traintest.py, [DATA_FILE (.csv)] is the output feature file extracted using the [ClassifyMode]CHARM-like.cppipe pipeline, and [OUTPUT_PATH] is the path to the directory where the output predicted_labels.csv will be created.

Notice

If images were acquired in several channels, either modify image loading in the CHARM-like.cppipe pipelines to handle multiple channels, or run the pipeline separately on either channels and concatenate the output files afterwards (custom scripts are available on demand for this).

Clone this wiki locally