KUCell
A dot-annotated cell detection dataset from Koç University, released publicly in 2026 — the nearest published analogue to a dataset Memorial could build itself.
What it is
KUCell was assembled at Koç University, Istanbul, and released alongside Çakı et al., Med Image Anal 2026;113:104155 as part of that paper’s contribution. It is distributed from the authors’ site at https://mysite.ku.edu.tr/cgunduz/downloads/KUCell.
What is known: it is an in-house cell-detection dataset used to evaluate a detection training strategy across multiple network architectures, alongside the public BRCA-M2C dataset as a comparator. The paper’s author list includes roughly eight pathologists and medical students across three Istanbul institutions — Koç University, Başakşehir Çam and Sakura City Hospital, and Şişli Hamidiye Etfal — which is consistent with a substantial manual annotation effort.
What is not known, because the full text is paywalled and the abstract does not say: the number of images, tissue types, stains, scanner, magnification, how many annotators marked each image, and whether any inter-annotator agreement was measured. [unverified] All of these would need checking before using it as a benchmark or comparing against local numbers.
Why it matters for my work
Two reasons, and the second is the more interesting.
As a benchmark, it is a public dot-annotated set that any local cell-detection work could be evaluated on, which converts an internal result into something comparable with published literature — the same argument made under External Validation.
As a precedent, it is a pathology department releasing its own annotated dataset, from a department a short distance away with a similar case mix and the same regulatory environment. Cell Detection and Counting notes that dot annotation is cheap enough to generate during routine work, which makes an in-house equivalent realistic in a way that boundary-annotated or slide-level-labelled datasets are not. KUCell is the worked example of that being done and published.
How it connects
Cell Detection and Counting — the task this dataset exists to support, and the annotation format it uses.
Çakı et al. 2026 — the paper that released it, and the only documentation of it currently in this wiki.
External Validation — a public dataset from another institution is exactly the kind of cohort that turns an internal result into a transferable one.
Memorial WSI Share — the local counterpart; if an in-house dot-annotated set were ever built, this is where the images would come from and the same de-identification rules would apply.
Interobserver Agreement — whether multiple annotators marked the same images, and how much they agreed, is the missing number that would determine how good a reference standard this is.
pT1-HBTG — the second nearby-group data release filed here (Bern/Radboud, 2026-08-23), and the instructive contrast: KUCell’s dot annotations are human and cheap, pT1-HBTG’s labels are outcomes — expensive because they need years of follow-up — while its cell positions are CNN detections rather than annotations. Together they bracket what “releasing a dataset” can mean.
Open questions
- Download it and record the basics: image count, tissue and stain, magnification, scanner, annotation protocol. That single action would close most of the
[unverified]tags above and is the obvious next step. - Is there any inter-annotator agreement figure? With that many pathologist authors it would be a natural thing to have measured, and its absence from the abstract does not mean absence from the paper.
- BRCA-M2C, the public comparator used in the same paper, has no page here yet. Worth one if the group starts benchmarking seriously.