Machine learning

CellDetector can detect cells (or other objects) in microscopy images such as histopathology, fluorescence, phase contrast, bright field, etc. It uses a machine learning-based method where a cell model is learned from simple dot annotations on a few images for training and predict on test sets. The installation requires some efforts but the instruction is well explained. Training parameters should be tuned for different datasets, but the default settings could be a good starting point.

This workflow is used to track multiple (appear/disappear, dividing and merging) objects in presumably big 2D+t or 3D+t datasets. It is best suitable for roundish objects or spots. Tracking is done through segmentation, which can be obtained from ilastik pixel classification, or imported from other tools. Users should provide a few object level labels, and the software predicts results on the rest of the image or new images with similar image characteristics. As a result, all objects get assigned random IDs at the first frame of the image sequence and all descendants in the same track (also children objects such as daughter cells) inherit this ID.

This macro batch processes all the 2D images (tif and jpg files) located in a user defined folder by calling Fiji Weka trainable segmentation to classify each pixel, and reports the areas of each class in a human readable results table. The classifier to be applied to each image should be previously trained on a representative image by an expert and exported to file (Save classifier) into the image folder to be processed.

ilastik is a simple, user-friendly tool for interactive image classification, segmentation and analysis. It is built as a modular software framework, which currently has workflows for automated (supervised) pixel- and object-level classification, automated and semi-automated object tracking, semi-automated segmentation and object counting without detection. Most analysis operations are performed lazily, which enables targeted interactive processing of data subvolumes, followed by complete volume analysis in offline batch mode. Using it requires no experience in image processing.

ilastik (the image learning, analysis, and segmentation toolkit) provides non-experts with a menu of pre-built image analysis workflows. ilastik handles data of up to five dimensions (time, 3D space, and spectral dimension). Its workflows provide an interactive experience to give the user immediate feedback on the quality of the results yielded by her chosen parameters and/or labelings.

The most commonly used workflow is pixel classification, which requires very little parameter tuning and instead offers a machine learning technique for segmenting an image based on local image features computed for each pixel.

Other workflows include:

Object classification: Similar to pixel classification, but classifies previously segmented objects by object characteristics in a subsequent step

Autocontext: This workflow improves the pixel classification workflow by running it in multiple stages and showing each pixel the results of the previous stage.

Carving: Semi-automated segmentation of 3D objects (e.g. neurons) based on user-provided seeds

Manual Tracking: Semi-automated cell tracking of 2D+time or 3D+time images based on manual annotations

Automated tracking: Fully-automated cell tracking of 2D+time or 3D+time images with some parameter tuning

Density Counting: Learned cell population counting based on interactively provided user annotation

Strengths: interactive, simple interface (for non-experts), few parameters, larger-than-RAM data, multi-dimensional data (time, 3D space, channel), headless operation, batch mode, parallelized computation, open source

Weaknesses: Pre-built workflows (not reconfigurable), no plugin system, visualization sometimes buggy, must import 3D data to HDF5, tracking requires an external CPLEX installation

Supported Formats: hdf5, tiff, jpeg, png, bmp, pnm, gif, hdr, exr, sif

This workflow estimates (densely distributed) object counts by the density of objects in the image without performing segmentation or object detection. Current version only works for 2D images of roundish objects with similar sizes on relatively homogeneous background. Users should provide a few labels of background and objects (especially on clustered objects), and the tool predicts the density of objects on the entire image. Counting is then estimated by integrating the density values on the whole image or specified rectangular regions of interests.