Implementation and complexity of the watershed-from-markers algorithm computed as a minimal cost forest

TL;DR

This paper presents an optimized implementation of the watershed-from-markers algorithm using Image Foresting Transform, significantly reducing memory usage while maintaining efficiency for large 3D datasets.

Contribution

It introduces five implementation modifications of the IFT algorithm, achieving minimal memory consumption suitable for large 3D medical images.

Findings

Memory usage reduced to 19-45% of typical datasets

Maintains O(m+C) time complexity

Enables segmentation of large 3D datasets on standard PCs

Abstract

The watershed algorithm belongs to classical algorithms in mathematical morphology. Lotufo et al. published a principle of the watershed computation by means of an Image Foresting Transform (IFT), which computes a shortest path forest from given markers. The algorithm itself was described for a 2D case (image) without a detailed discussion of its computation and memory demands for real datasets. As IFT cleverly solves the problem of plateaus and as it gives precise results when thin objects have to be segmented, it is obvious to use this algorithm for 3D datasets taking in mind the minimizing of a higher memory consumption for the 3D case without loosing low asymptotical time complexity of O(m+C) (and also the real computation speed). The main goal of this paper is an implementation of the IFT algorithm with a priority queue with buckets and careful tuning of this implementation to reach as minimal memory consumption as possible. The paper presents five possible modifications and methods of implementation of the IFT algorithm. All presented implementations keep the time complexity of the standard priority queue with buckets but the best one minimizes the costly memory allocation and needs only 19-45% of memory for typical 3D medical imaging datasets. Memory saving was reached by an IFT algorithm simplification, which stores more elements in temporary structures but these elements are simpler and thus need less memory. The best presented modification allows segmentation of large 3D medical datasets (up to 512x512x680 voxels) with 12-or 16-bits per voxel on currently available PC based workstations.