[Paper Review] Efficient Column Generation for Cell Detection and Segmentation

Chong Zhang, Shaofei Wang+2 more|arXiv (Cornell University)|Sep 21, 2017

Advanced Neural Network Applications44 references3 citations

TL;DR

This paper proposes a novel column generation approach for cell instance segmentation in crowded biological images, formulating the problem as an integer program with non-overlapping cell constraints. By leveraging superpixel-based set packing and odd set inequalities to tighten the LP relaxation, the method achieves state-of-the-art accuracy with fast, anytime inference—demonstrated on three microscopy datasets with near-optimal gaps and sub-minute inference times on average.

ABSTRACT

We study the problem of instance segmentation in biological images with crowded and compact cells. We formulate this task as an integer program where variables correspond to cells and constraints enforce that cells do not overlap. To solve this integer program, we propose a column generation formulation where the pricing program is solved via exact optimization of very small scale integer programs. Column generation is tightened using odd set inequalities which fit elegantly into pricing problem optimization. Our column generation approach achieves fast stable anytime inference for our instance segmentation problems. We demonstrate on three distinct light microscopy datasets, with several hundred cells each, that our proposed algorithm rapidly achieves or exceeds state of the art accuracy.

Motivation & Objective

• Address the challenge of accurate and efficient instance segmentation in densely packed, biologically complex microscopy images with minimal human intervention.
• Overcome the computational intractability of enumerating all possible cell instances by formulating instance segmentation as a large-scale set packing problem.
• Enable stable, anytime inference for real-time or high-throughput bioimage analysis by leveraging column generation with exact pricing on small-scale IPs.
• Improve the quality of the LP relaxation without requiring branch-and-price by integrating odd set inequalities directly into the pricing phase.
• Demonstrate robustness across diverse cell shapes and imaging modalities, particularly in transparent or low-contrast images.

Proposed method

• Formulate instance segmentation as an integer linear program (ILP) where variables represent candidate cells and constraints enforce pixel-wise non-overlap via set packing.
• Use superpixels as atomic image units to reduce dimensionality and enable tractable optimization, with each superpixel representing a potential element in the set packing formulation.
• Implement a column generation framework where the pricing subproblem is solved via exact optimization of small-scale integer programs, enabling efficient column generation.
• Tighten the LP relaxation using odd set inequalities that are compatible with the pricing problem structure, avoiding the need for branch-and-price.
• Apply dual ascent and Lagrange relaxation techniques to compute lower bounds and guide the column generation process.
• Integrate a random forest classifier to estimate unary potentials for superpixels based on intensity and boundary features, which inform the cost function in the ILP.

Experimental results

Research questions

• RQ1 Can a column generation approach with exact pricing on small-scale IPs achieve high-quality instance segmentation in crowded biological images with minimal training?
• RQ2 How effective is the integration of odd set inequalities in tightening the LP relaxation without disrupting the pricing problem structure?
• RQ3 To what extent can superpixel-based dimensionality reduction preserve segmentation accuracy while enabling scalable optimization?
• RQ4 How does the proposed method compare in accuracy and inference speed to state-of-the-art methods on real microscopy datasets with varying cell morphology?
• RQ5 Can the method provide stable, anytime solutions with tight duality gaps across diverse biological image datasets?

Key findings

• The method achieved or exceeded state-of-the-art performance on three distinct light microscopy datasets, with F-scores above 0.90 and Dice coefficients exceeding 0.85 in most cases.
• For Dataset 1, 99.28% of instances achieved a normalized duality gap below 0.1, indicating high solution quality and tight relaxation.
• Inference times were typically under 200 seconds (peak at ~150s for Dataset 1), with the method being approximately an order of magnitude faster than the prior state-of-the-art (Zhang et al., 2014b).
• The approach required minimal human annotation, as it relied on a single random forest classifier per dataset with no need for fully labeled training data.
• The method demonstrated strong robustness to cell shape and size variation, successfully segmenting highly non-circular and overlapping cells in the same image.
• The use of odd set inequalities improved the LP relaxation without breaking pricing problem structure, enabling efficient solution without branch-and-price.