Quality Aware Candidate Selection for Silver Truth in Cell Segmentation
Abstract Background.: Quantitative microscopy requires accurate cell masks, but cell segmentation accuracy varies across individual cells and across segmentation models. Combining candidate masks available for a cell can yield more accurate segmentations, especially when poor-quality candidates are identified and eliminated. We test whether instance-level quality ranking and selection of candidate masks improves label fusion accuracy. Methods.: We evaluate segmentation on five Cell Tracking Challenge datasets. Using reference tracking markers to collect candidate masks from competitor submissions for each cell, we extract aligned crops of the candidate masks and the underlying microscopy image. A deep quality estimator predicts the intersection over union (IoU) for each candidate, and a pairwise ranker prunes degraded masks to retain the most promising candidates. Our ENSEMBLE pipeline combines the retained candidates using learned Deep Sets weighting and a boundary-aware vote. For a controlled comparison, we also apply the established SIMPLE fusion method to the same selected candidates as well as unfiltered pools. Using a two-fold cross-validation scheme across paired video sequences, we evaluate IoU on aligned cell crops and on fully reconstructed segmentation images. Results.: The Deep Sets ENSEMBLE achieves 0.8676 mean IoU. SIMPLE with the same quality-based candidate selection achieves 0.8654, compared with 0.8431 for the descriptive best fusion baseline without candidate selection. Reconstructing Deep Sets predictions as full images yields 0.8664 cell-weighted mean IoU with zero cell loss. The small difference between ENSEMBLE and SIMPLE indicates that instance-level candidate selection produces most of the improvement. The descriptive reference formed by selecting the best individual competitor in each fold averages 0.8712. Conclusions.: Ranking and pruning candidate masks for each individual cell substantially improves label fusion when candidate quality varies. ENSEMBLE achieves the leading fusion result, while the comparable performance of Selected SIMPLE demonstrates that the primary practical gain stems from quality-aware candidate selection prior to fusion.
Authors
- Monika Weissová
- Vladimír Ulman (ORCID: https://orcid.org/0000-0002-4270-7982)
- David Číž
- João Manuel Da Silva Carias
Institutions
- VSB - Technical University of Ostrava (CZ)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11515932/v1
- Primary Topic
- Cell Image Analysis Techniques
- Type
- preprint
Funders
- NextGenerationEU