Bridging 3D Deep Learning and Curation for Analysis and High-Quality Segmentation in Practice

TL;DR

This work presents VessQC, an open-source, uncertainty-guided 3D segmentation curation tool implemented as a napari plugin to streamline human-in-the-loop quality control. By integrating uncertainty maps with topology-aware cues, it prioritizes edits to the most informative regions and supports interactive 3D correction. In a preliminary study on $100^3$-voxel vascular volumes, uncertainty-guided curation markedly improved error-detection recall ($p=0.007$) with no significant per-correction slowdown, albeit with a modest total time increase ($p=0.084$). The approach demonstrates a practical, data-centric workflow for enhancing 3D bioimage segmentation, adaptable to a broad range of 3D datasets beyond vasculature.

Abstract

Accurate 3D microscopy image segmentation is critical for quantitative bioimage analysis but even state-of-the-art foundation models yield error-prone results. Therefore, manual curation is still widely used for either preparing high-quality training data or fixing errors before analysis. We present VessQC, an open-source tool for uncertainty-guided curation of large 3D microscopy segmentations. By integrating uncertainty maps, VessQC directs user attention to regions most likely containing biologically meaningful errors. In a preliminary user study uncertainty-guided correction significantly improved error detection recall from 67% to 94.0% (p=0.007) without a significant increase in total curation time. VessQC thus enables efficient, human-in-the-loop refinement of volumetric segmentations and bridges a key gap in real-world applications between uncertainty estimation and practical human-computer interaction. The software is freely available at github.com/MMV-Lab/VessQC.

Figures (2)

• Figure 1: VessQC Interactive Curation Interface (Napari). The interface displays a selected region extracted from the full 3D segmentation with the chosen branch visually highlighted. User tools for image modification (e.g., brightness, contrast) and view control (2D/3D toggle) are provided. The highlighted branch can be directly edited in 2D. Once edits are complete, changes are saved via the “Done” button, and the next branch for review can be immediately selected from the segmentation list on the right.
• Figure 2: Curation workflow in VessQC. A 3D overview of the vascular sample is shown (a), with the highlighted region containing a vessel flagged for high topology-based uncertainty ($\mathbin{\approx}0.8$). The pre-curation segmentation (red, in b) reveals a potential false merge between neighboring vessels, as indicated by the uncertainty cue (yellow). Inspection of the corresponding raw image data (c) confirms that the vessels are not physically connected. After manual correction within VessQC, the erroneous connection is removed and the two vessels are correctly separated (d).