An open-source, MATLAB based annotation tool for virtual slides
Riku Turkki, Margarita Walliander, Ville Ojansivu, Nina Linder, Mikael Lundin, Johan E. Lundin · Diagnostic Pathology · 2013
Computer aided analysis of virtual slide images has become an increasingly popular research topic and area of development. Novel applications are aimed for instance at automation of certain stages in sample assessment and assay readout (e.g. segmentation, detection, etc.) [ 2 ] or for reproducible measurement of a sample’s visual appearance (e.g. grading, morphological classes etc.) [ 3 ]. Statistical learning methods are one of the key algorithms used in building these image analysis applications. Especially supervised algorithms (e.g. support vector machines (SVM), AdaBoost, k nearest neighbour (kNN), etc.) offer a way to train a classifier that can perform complex quantification tasks. The process requires annotation of labelled examples in order to describe the task and learn a model. Recently, image acquisition techniques for large-scale digitization of tissue samples have become common and require new methods to perform the annotation and associated labelling [ 4 ]. In this study we present an annotation tool that combines 1) direct interaction with a remote slide collection (a web based virtual microscopy application), 2) an interface to annotate points of interest in the virtual slides and 3) fast transition from annotation to development of image analysis methods. The annotation tool is written in MATLAB (matrix laboratory) that is a cross-platform numerical computing environment (MathWorks, Natick, MA). MATLAB is a tool with a wide range of application areas and adopted by the image analysis community because of its versatile features. The programming language and its computing environment offers tools for testing and implementation of methods from a number manipulation to creating interactive graphical user interface (GUI) applications. The annotation tool is based on exploiting two important properties of the current platform: 1) a random code-stream access featured image format (e.g. ECW or JPEG 2000) and 2) a compression streamlining protocol. The random code-stream compression makes it possible to extract sub-images from a large virtual slide file. Thus, the user can download only a field of interest, instead of loading the whole image file, which might be of gigapixel size. The other required property, the compression streamlining protocol, basically implements this sub-image extraction over standard hypertext transfer protocol (HTTP) in an efficient way. With the advantages of the above-mentioned properties that a slide collection should have, it is reasonably straightforward to implement a MATLAB based tool to operate a remote database. The overall structure of the tool is illustrated in Figure 1 . In a first phase the tool reads the metadata of a virtual slide file and deduces its height and width. Using a server’s streamlining protocol, uniform resource locators (URL) are defined to extract the tiles from the pre-calculated coordinates within the virtual slide. In the annotation phase, the tiles are extracted and downloaded one at the time from the server and displayed to an annotator and requested for annotations. The locations and labels of the annotations are saved locally in MAT files. The operational principal of the annotation tool A) A slide collection saved at a remote virtual microscopy server, B) the tool reads the metadata of the slide and determines the coordinates for the tiles, C) a tile is downloaded and displayed to an annotator and the coordinate-label pairs are saved to a client computer.