Answers are grounded in this paper and cite their sources.
The detection and counting of cell nuclei in histological images remain difficult because of spectral variability and overlapping structures. This preliminary study investigates whether spectral band selection, combined with ellipse-based geometric validation, can improve classical unsupervised nuclei detection and counting in multispectral histology images. Three classical algorithms (Otsu thresholding, K-means clustering and the Expectation–Maximization (EM) algorithm) were applied to every band of four multispectral configurations (3, 5, 10 and 25 bands) acquired from the same rat brain tissue specimen. An ellipse-fitting step was used to reject non-elliptical artifacts and to flag contours containing several nuclei. All results were produced by an open-source Python implementation applied to the public dataset. Detection was evaluated with a nucleus-level overlap rule, and counting with a one-to-one matching rule that penalizes both merged and split nuclei; in addition, estimated and reference numbers of nuclei were compared region by region. Without ellipse fitting, the best configurations reached F-scores of 91.95% for detection (EM, 10-band acquisition, band 6) and 80.41% for counting (EM, 10-band acquisition, band 5), and the smallest mean absolute count error was 7.0 nuclei per region (about 5% of the reference count; K-means, 10-band acquisition, band 5). Mid-spectrum bands consistently outperformed extreme bands, and band choice had a larger effect than the choice of algorithm. Ellipse fitting increased detection precision (by 0.5–9.9 percentage points) at the cost of detection recall, whereas its multiple-nucleus rule inflated the estimated number of nuclei and lowered the counting F-score in all band/algorithm combinations. Because all spectral views originate from a single specimen and the best bands were selected on the evaluated data, these results are exploratory, apply only to the specimen studied, and must be confirmed on independent specimens, tissue types and staining protocols.
No takes yet. Share an insight, caveat, or question.
Boudabous et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: