CenSegNet

Centrosome abnormalities — extra centrosomes (numerical) or enlarged, malformed ones (structural) — are a long-standing hallmark of epithelial cancer, first linked to tumorigenesis by Theodor Boveri more than a century ago. Quantifying them has stayed stubbornly manual: centrosomes are sub-micron puncta scattered through dense, layered tissue, and generic pixel-wise segmentation networks either miss them or hallucinate them in cluttered regions. That bottleneck has kept centrosome biology confined to cell culture and small tissue series.

CenSegNet, developed at the University of Southampton with collaborators in Cambridge, is a modular deep learning framework that automates centrosome detection, morphometry, and per-cell assignment directly in intact tissue. Rather than segmenting whole slides in one pass, it decomposes the problem into detection, patch-level refinement, and epithelial cell delineation, so that each stage operates at the scale its target demands. The work was posted as a preprint in September 2025 and published in Nature Communications in July 2026.

Where generalist cell segmenters such as Cellpose and CellSAM resolve cells and nuclei, CenSegNet targets an organelle an order of magnitude smaller and couples it to cell-level assignment — turning a qualitative pathology observation into a per-cell, spatially indexed measurement.

#Key Features

• Three-stage modular pipeline: A YOLOv11-seg detector proposes centrosome candidates, a U-Net refines each candidate's boundary, and StarDist delineates epithelial cells so that every centrosome can be assigned to a parent cell.
• Cross-modality operation: A single combined checkpoint handles brightfield immunohistochemistry, haematoxylin-stained sections, confocal immunofluorescence, and lightning-based super-resolution immunofluorescence, using a modality-specific input calibration step rather than separate models.
• Structural and numerical phenotyping: Per-centrosome area and shape descriptors separate structural abnormalities (above a 6.5 µm² threshold) from numerical amplification, which the paper shows are mechanistically uncoupled in breast tumours.
• Threshold robustness: F1, IoU, and Cohen's κ stay stable as detection and segmentation thresholds are swept from 0.1 to 0.8 across all four image types, so results are not tuned per dataset.
• GUI and CLI distribution: A PyQt5 desktop interface and a command-line entry point ship with the code, with pretrained weights and held-out example images for direct inference without retraining.

#Technical Details

The detector is a YOLOv11-seg model fine-tuned on over 15,000 expert-annotated centrosomes with a composite box, segmentation, classification, and distribution focal loss, trained with AdamW (learning rate 0.002, momentum 0.9, batch size 16) for 300 epochs. Detected candidates are cropped into 256×256-pixel centrosome-centred patches with 40-pixel padding and passed to a three-channel-in, one-channel-out U-Net trained for 100 epochs under BCEWithLogits loss with RMSprop (learning rate 1×10⁻⁴, weight decay 1×10⁻⁸, momentum 0.9). Cell boundaries come from StarDist's star-convex instance segmentation; on keratin 8-labelled validation tissue, its assignments tracked the marker at R² = 0.99 (immunofluorescence) and R² = 0.90 (immunohistochemistry). The implementation is Python 3.10 and PyTorch 2.1, and the annotated training corpus took over 200 person-hours to build in QuPath, with the immunofluorescence set annotated using Cellpose 3.

On held-out test data of 25 immunohistochemistry tissue microarray cores and 17 immunofluorescence images, predicted centrosome counts matched manual ground truth at R² = 0.9999 and R² = 0.9873 respectively. Mean F1 across modalities was 0.82, against 0.72 for U-Net, 0.68 for SegNet, and 0.65 for DeepLabv3+ trained under identical conditions on 108 annotated immunohistochemistry images (80 train / 8 validation / 20 test) plus a 25-image curated evaluation set.

#Applications

CenSegNet is aimed at pathology and cell-biology labs that need organelle-level readouts from tissue rather than from culture. In the source study it profiled 333,148 centrosomes across 911 breast tissue cores from 127 patients, resolving numerical abnormalities to proliferative tumour cores and structural abnormalities to invasive margins, and linking those spatial patterns to histological grade, hormone receptor status, HER2 expression, nodal involvement, and germline alterations. The authors also validated it on kidney, colon, and appendix epithelium, and the CLI supports batch processing of tissue microarray cohorts on a single GPU workstation.

#Impact

The framework converts centrosome amplification from a hand-counted descriptive trait into a scalable, spatially resolved biomarker, opening the door to correlating it with transcriptomic or proteomic profiles and to stratifying patients for PLK4, AURKA, and HSET inhibitors. Its scope is deliberately narrow — one organelle plus epithelial architecture, not general-purpose cell segmentation — and the survival association reported for structural abnormalities comes from a cohort the authors state is not powered for definitive prognostic claims. Code is BSD-3-Clause on GitHub with a Zenodo archive, but the combined pretrained checkpoint is distributed via a Google Drive link rather than a versioned model registry.