System and method for image segmentation

An image analysis system for segmenting a digital image of a biological sample, the digital image comprising at least a first image component depicting a first tissue type of the biological sample and a second image component depicting a second tissue type of the biological sample or background, the image analysis system comprising an interface for receiving the digital image and a processor configured for performing the segmentation, the segmentation comprising:

• generating a plurality of image channels by applying color deconvolution on the received digital image or by separating RGB color components of the digital image into respective image channels, wherein each of the image channels is a digital image whose pixel intensity values correspond to the color components of the respective channel;
• identifying a plurality of superpixels in the received digital image;
• for each of the superpixels and for each of the image channels, extracting a feature set, the feature set comprising and/or being derived from pixel intensity values of pixels contained in said superpixel;
• receiving at least a first and a second marking provided by a user or provided automatically, the first marking covering one or more first ones of the superpixels, the first marked superpixels representing regions of the first image component, the second marking covering one or more second ones of the superpixels, the second marked superpixels representing regions of the second image component;
• for each of the plurality of image channels, comparing the image channel specific feature set of at least one of the first marked superpixels with the image channel specific feature set of at least one of the second marked superpixels for obtaining an image channel specific marker-difference score;
• representing the plurality of superpixels as a graph, whereby the center of each of the identified superpixels is represented by a node and whereby the nodes representing centers of adjacent superpixels are connected by a respective edge, wherein the computation of each of the edges comprises:
• computing, for each of the image channels, a feature set distance by comparing the image channel specific feature sets of the neighboring superpixels connected by said edge;
• multiplying the feature set distance obtained for each of the edges and for each of the image channels with the image channel specific marker-difference score computed for said image channel, thereby computing image-channel specific edge weights;
• for each of the edges, aggregating all image-channel specific edge weights and using the aggregated edge weight as the edge weight of the edge;
• for each unmarked superpixel of the plurality of superpixels:
• computing a first combined distance between said unmarked superpixel and the one or more first marked superpixels by means of a graph traversal algorithm, the first combined distance being a derivative of a feature-set-dependent distance between said unmarked superpixel and the one or more first marked superpixels and of a spatial distance between said unmarked superpixel and the one or more first marked superpixels;
• computing a second combined distance between said unmarked superpixel and the one or more second marked superpixels by means of a graph traversal algorithm, the second combined distance being a derivative of a feature-set-dependent distance between said unmarked superpixel and the one or more second marked superpixels and of a spatial distance between said unmarked superpixel and the one or more second marked superpixels;
• assigning the unmarked superpixel to the first image component if the first combined distance is smaller than the second combined distance and otherwise associating the unmarked superpixel to the second image component, thereby segmenting the digital image.