Deep MIB - Directories and Preprocessing Tab

Back to MIB | User interface | DeepMIB

Configuration of directories and preprocessing settings for deep learning segmentation in Microscopy Image Browser.

Overview

The Directories and Preprocessing tab in Deep MIB allows users to specify directories containing images for training and prediction, along with parameters for image loading and preprocessing.

Widgets and settings

Directories

Directory with images and labels for training

Training directory panel

Used only for training

Select the directory containing images and models for training (named 1_Training in directory organization schemes below).
For 2D networks, use individual 2D images; for 3D networks, use individual 3D datasets.

extension: specifies the image file extension
Bio: toggles between standard and Bio-Formats readers. For Bio-Formats collections, use Index... to specify
the file index within the container

For better performance, convert Bio-Formats images to standard formats or use preprocessing (see below).

important notes considering training files

Number of model/mask files must match the number of image files, except for 2D networks where a single .model file is allowed if Single MIB model file is checked (requires preprocessing)
For standard image format labels, specify the total number of classes (including Exterior) in Number of classes
Important: avoid numeric material names in .model format; use descriptive names instead

Directory with images for prediction

Prediction directory panel

Used only for prediction

Specify the directory with images for prediction (named 2_Prediction in directory organization schemes below).
Place images in an Images subfolder (or directly in the specified folder). Optionally, include ground truth labels in a Labels subfolder.

Info

In preprocessing mode, images are converted and saved to 3_Results/PredictionImages.
Fround truth labels, if present, are processed to 3_Results/PredictionImages/GroundTruthLabels for evaluation (see Predict tab).
For 2D networks, use 2D images or 3D stacks; for 3D networks, use 3D datasets.

extension: specifies the image file extension

Bio: toggles between standard and Bio-Formats readers. For Bio-Formats collections, use Index to specify the file index.

Directory with resulting images

Results directory panel

Specify the main output directory where results and preprocessed images are stored. Deep MIB automatically creates subfolders:

PredictionImages: preprocessed images for prediction
PredictionImages/GroundTruthLabels: ground truth labels for prediction images, if available
PredictionImages/ResultsModels: main output directory for generated labels after prediction. Combine 2D models in MIB using Shift + during loading
PredictionImages/ResultsScores: prediction scores (probability) for each material, scaled 0-255
ScoreNetwork: accuracy/loss plots (if Export training plots is enabled in the Train tab) and network checkpoints (if Save progress after each epoch is checked).
Scores are timestamped and overwritten with new training
TrainImages: preprocessed training images (preprocessing mode only)
TrainLabels: preprocessed training labels (preprocessing mode only)
ValidationImages: preprocessed validation images (preprocessing mode only)
ValidationLabels: preprocessed validation labels (preprocessing mode only)

Label file details

Label file details panel

Single MIB model file: (2D networks only) uses a single .model file for labels
Labels extension: (2D networks only) selects the model file extension. 3D networks use MIB .model format
Number of classes: (TIF/PNG only) defines the number of classes, including Exterior. Auto-updated for .model files
Use masking: excludes parts of training data using masks (file count must match images). with preprocessing, requires .mask format; without, use USE 0-s IN LABELS for 0-value areas in labels (recommended to skip preprocessing).

Info

With USE 0-s IN LABELS, the first predicted material is Exterior (index 0). assign the first ground truth material to background
With preprocessing masks, Exterior indicates background
Masking may reduce precision due to patch inconsistency; minimize its use

Mask extension: selects mask file extension. With preprocessing, only .mask is allowed for 3D networks; without, any format is permitted

Additional settings

Additional settings panel

Compress processed images compresses preprocessed images to .mibImg format (loadable in MIB or MATLAB, e.g., res = load('img01.mibImg', '-mat');). Slows down performance
Compress processed labels compresses preprocessed labels to .mibCat format (loadable via Menu → Models → Load model). Slows performance but reduces file size significantly
Use parallel processing enables multi-core processing, with core count set in Workers. Speeds up preprocessing
Fraction of images for validation sets the fraction of images randomly assigned to validation (based on Random generator seed). If 0, validation is skipped
Random generator seed initializes the random seed for splitting training/validation images. fixed values ensure reproducibility; If 0 uses system time for randomness
Preprocess for selects the mode for the Preprocess button (see schemes below)

Preprocessing of files

Preprocessing of files was originally required for most workflows, however now Deep MIB supports unprocessed images in many cases.
In this case the Preprocess for should be set to

Split for training/validation to automatically split images into the training and validation sets
Preprocessing is not required after images were split into the training and validation sets

When the preprocessing step is required or recommended

Preprocessing is recommended or required when:
- Labels are in a single .MODEL file (for 2D workflows)
- Training data uses proprietary formats readable only by Bio-Formats
During preprocessing, images and models are converted to .mibImg and .mibCat formats (MATLAB-based) optimized for training and prediction.

Organization of directories

Sections below provide schemes for directories organization for several cases:

Automatic file splitting [recommended for most cases] without file conversion, with automatic splitting of files for training and validation
Manual file splitting without file conversion, files arranged manually into correct folders
Conversion of files with preprocessing/file conversion, files are converted to .mibImg and .mibCat formats and split for training and validation
Patch-wise workflow this workflow requires slightly different organization of directories

Automatic file splitting

Without file conversion, with automatic splitting of datasets into training and validation sets.
Images and labels are randomly split into training and validation sets upon clicking
Preprocess when Preprocess for: Split for training and validation.
Splitting depends on the Random generator seed value (when "0" uses a new random seed each time).

Directory tree for automatic file splitting without file conversion

Snapshot with the legend

Manual file splitting

Without file conversion, when datasets are manually split into training and validation sets.
Images for training are loaded on-demand without preprocessing.

In order to proceed, split files into TrainImages, TrainLabels, and optional ValidationImages, ValidationLabels subfolders (see Snapshot with the directory tree below).
Automatic splitting is also available (see Automatic file splitting). For Bio-Formats, preprocessing is recommended to improve file reading speed.

Directory tree for manual file splitting without file conversion

Snapshot with the legend

Conversion of files

Organization of directories with file conversion during preprocessing and automatic split for training and validation.

Enabled when Preprocess for is set to:
- Training and prediction: preprocesses for both
- Training: preprocesses only training images
- Prediction: preprocesses only prediction images

Start conversion by pressing Preprocess.

Directory tree for conversion of files

Snapshot with the legend

Patch-wise workflow

Organization of directories for 2D patch-wise workflow. The 2D patch-wise workflow organizes training images in Images/[ClassnameN] subfolders (e.g., bg, spots).
Number of folders should match the number of classes.

Patch-wise, manually split files into training and validation sets

Place training images in subfolders named by class under:
- 1_Training/TrainImages: training images
- 1_Training/ValidationImages: validation images (optional)
The images can be automatic split (see the next section).

For prediction ground truth, use 2_Prediction/Images and 2_Prediction/Labels (semantic style) or 2_Prediction/[ClassnameN] (patch-wise style).

Directory tree for patch-wise, manually split files

bg and spots are example class names.

Snapshot with the legend

Patch-wise, automatic split files into training and validation sets

Automatic splitting of datasets for the patch-wise segmentation into training and validation sets

Images are randomly split (based on Random generator seed) into training and validation sets upon clicking
Preprocess when Preprocess for: Split for training and validation.
Initially, place all images in 1_Training/Images/[ClassnameN].
Prediction ground truth follows the same options as above.

Directory tree for patch-wise, automatic split files

bg and spots are example class names.

snapshot with the legend

Back to MIB | User interface | DeepMIB