Efficient Interactive Multi-object Segmentation in Medical Images

In medical image segmentation, it is common to have several complex objects that are difficult to detect with simple models without user interaction. Hence, interactive graph-based methods are commonly used in this task, where the image is modeled as a connected graph, since graphs can naturally represent the objects and their relationships. In this work, we propose an efficient method for the multiple object segmentation of medical images. For each object, the method constructs an associated weighted digraph of superpixels, attending its individual high-level priors. Then, all individual digraphs are integrated into a hierarchical graph, considering structural relations of inclusion and exclusion. Finally, a single energy optimization is performed in the hierarchical weighted digraph satisfying all the constraints and leading to globally optimal results. The experimental evaluation on 2D medical images indicates promising results comparable to the state-of-the-art methods, with low computational complexity.