NuSeC: Breast NST Nuclei Segmentation

• NuSeC is a dedicated dataset providing manually annotated H&E images of invasive breast carcinoma NST for rigorous nuclei segmentation benchmarking.
• It employs a strict patient-level split from 25 patients (100 images) at 40× magnification to ensure reliable comparative analysis of CAD systems.
• The dataset emphasizes instance-level evaluation using metrics like AJI and IoU, guiding methodological best practices for post-processing and nuclei separation.

NuSeC is a publicly available dataset designed specifically for nuclei segmentation in hematoxylin and eosin (H&E) stained breast cancer histopathology images. Its stated goal is to support the development and comparative evaluation of computer-aided diagnosis (CAD) systems for robust and reliable segmentation of nuclear structures in invasive breast carcinoma, facilitating downstream diagnostic tasks. The dataset is defined by a single-tissue focus on invasive breast carcinoma no special type (NST), a strict patient-level split for benchmarking, and manually annotated nuclei masks acquired at high magnification, with evaluation centered on Aggregated Jaccard Index (AJI) and Intersection over Union (IoU) (Samet et al., 18 Jul 2025).

1. Domain definition and research scope

NuSeC is scoped to invasive breast carcinoma NST and therefore occupies a narrower domain than multi-organ nuclei segmentation resources. The paper identifies this single-tissue focus as a distinctive aspect of the dataset, alongside patient-aware splitting and manual annotation. All images are H&E stained breast cancer histopathology images, and the intended use cases are training and benchmarking nuclei segmentation models, method comparison under a consistent evaluation protocol, and research on H&E variability and model robustness in breast cancer histopathology (Samet et al., 18 Jul 2025).

This single-tissue design has methodological consequences. Because the tissue type and pathology context are fixed to breast cancer, specifically invasive carcinoma NST, the benchmark is coherent for breast pathology research and model development. This suggests a dataset oriented toward controlled comparative analysis within a defined breast pathology setting rather than toward broad, cross-tissue generalization. The paper correspondingly notes that models trained on NuSeC may need additional data to generalize to other tissues or breast cancer subtypes.

A further point of scope is institutional provenance. The material is sourced from the Department of Medical Pathology at Ankara University. The paper does not discuss ethical or IRB approvals and consent.

2. Data acquisition and corpus composition

NuSeC is constructed from 25 different invasive breast carcinoma (NST) patients. From the slides of each patient, 4 images with the size of 1024×1024 pixels were selected, yielding 100 images in total. The images were captured at 40× magnification, the slides were scanned by a 3D Histech Panoramic P250 Flash-3 scanner, and an Olympus BX50 microscope was used in the imaging pipeline. The stain is H&E, described in the paper as the routine stain used globally in pathology to provide clear contrast of nuclei and cytoplasm (Samet et al., 18 Jul 2025).

The paper reports approximately 36,000 nuclei across the dataset. Quantitative distribution characteristics beyond the train/test totals are not provided: per-image nuclei counts, min/median/max statistics, and quantitative variability across patients are not reported. Pixel size in $\mu m$/pixel is also not reported.

Component	Count	Notes
Patients	25	Invasive breast carcinoma (NST)
Images per patient	4	Selected from each patient’s slide(s)
Total images	100	Each image is 1024×1024
Training set	75 images	Around 30,000 nuclei structures
Test set	25 images	Around 6,000 nuclei structures

Because all images are H&E stained and acquired through the same named scanner and microscope pipeline, the dataset is more controlled than heterogeneous multi-institutional whole-slide image collections. At the same time, the paper notes that acquisition variability may still arise from the use of the 3D Histech P250 Flash-3 scanner and Olympus BX50 microscope, although stain or color distribution variability is not quantified.

3. Split protocol and annotation model

The official split strategy is 75%/25% at the patient level. For each of the 25 patients, 1 of the 4 images was randomly selected into the test set, and the remaining 3 images were allocated to the training set. This yields 75 training images and 25 test images. The paper explicitly frames this strategy as a mechanism for consistent comparative analysis across future methods and as a safeguard against leakage across patients (Samet et al., 18 Jul 2025).

The validation split is not specified in the paper. The recommended practice is to create a validation set from the training portion while ensuring that it remains patient-disjoint from the test set. This is a central operational constraint for any experimental protocol built on NuSeC.

The annotation of nuclear structures was carried out manually using QuPath, and a mask image was created for each image. Although the paper does not specify the file format, naming conventions, metadata layout, annotation guidelines, or explicit boundary definitions, it states that the evaluation is instance-centric through AJI. That implies that the ground truth comprises per-nucleus segments suitable for instance-level evaluation. The paper further notes that pathologists from Ankara University are listed as co-authors and that annotation was performed within a pathology context, but it does not provide specific details on annotator roles, quality assurance procedures, or inter-annotator agreement statistics.

A common source of confusion in reuse is to treat NuSeC as if it provided a generic patch-level or slide-level split. It does not. The benchmark is defined around image-level samples drawn from patients under a patient-level train/test partition, and the methodological requirement is that images from the same patient do not cross between train/validation and test subsets.

4. Evaluation protocol and official metrics

The paper specifies two evaluation metrics: AJI and IoU. AJI is the more structurally informative of the two because it operates on nucleus instances rather than only on foreground masks. Let the set of ground-truth nucleus instances be $G = \{G_i\}_{i=1}^N$, and the set of predicted instances be $P = \{P_k\}_{k=1}^M$. For each ground-truth instance $G_i$, its matched prediction $P_{m(i)}$ is the predicted instance that yields the largest Jaccard index with $G_i$. Let $U$ be the set of unmatched predicted instances. The Aggregated Jaccard Index is defined as

$$\mathrm{AJI} = \frac{\sum_{i=1}^{N} \left| G_i \cap P_{m(i)} \right|}{\sum_{i=1}^{N} \left| G_i \cup P_{m(i)} \right| + \sum_{k \in U} \left| P_k \right|}.$$

As described in the paper, AJI ranges from 0 to 1, with higher values indicating better agreement (Samet et al., 18 Jul 2025).

IoU is the second official metric. The paper describes it as measuring the intersection over union of nuclei areas on a ground-truth mask and defines it in terms of TP, FP, and FN. No other official metrics are reported. In particular, Dice, pixel-level F1, boundary F1, and Panoptic Quality are not part of the stated protocol.

The paper does not report baseline architectures, training settings, or quantitative results on the official test set. Consequently, NuSeC is presented as a benchmark substrate rather than as a leaderboard paper. Community benchmarks therefore need to be established by later work, and results should be reported on the official patient-disjoint test set using AJI and IoU.

The practical implication is that post-processing matters. The paper advises that instance separation may be required if a network outputs semantic masks, for example via watershed or boundary-aware methods, to align with AJI’s instance-level evaluation. It also cautions that post-processing should not merge nuclei improperly.

5. Relation to adjacent nuclei-segmentation datasets

The paper situates NuSeC relative to several existing datasets. Compared with MoNuSeg, described as a multi-organ nucleus segmentation resource/challenge, NuSeC offers a breast-specific benchmark with patient-level splitting and 40× magnification images at 1024×1024 resolution. Compared with TNBC, which focuses on triple-negative breast cancer, NuSeC covers invasive breast carcinoma NST. CoNSeP is identified as a colon epithelial nuclei dataset and therefore differs in tissue type and pathology context. CryoNuSeg is described as a cryosection-based nuclei segmentation dataset and differs in specimen preparation workflow compared to paraffin-embedded, H&E-stained slides in NuSeC (Samet et al., 18 Jul 2025).

These comparisons clarify what NuSeC is not. It is neither a multi-organ benchmark nor a general-purpose breast cancer nuclei resource spanning multiple breast cancer subtypes. Rather, it complements multi-organ and other single-organ datasets by providing a focused breast cancer NST collection with explicit patient-level splits and AJI/IoU evaluation guidance.

A plausible implication is that NuSeC is especially useful when the research question emphasizes within-domain benchmarking, stain variability under a fixed pathology context, or breast NST-specific segmentation behavior. Conversely, it is less suitable as a sole training corpus for claims about broad tissue generalization.

6. Practical use, access, and limitations

The paper does not prescribe preprocessing steps. It instead notes that, in typical practice for H&E nuclei segmentation on 40× images, users may consider color normalization, patch tiling or cropping, intensity normalization, stain augmentation, and standard data augmentation such as flips, rotations, elastic deformations, and color jitter. Since images are already 1024×1024, further tiling may be optional depending on GPU memory and model input size. Any additional internal validation split should respect patient disjointness, and stain normalization parameters should be learned exclusively from training data to avoid leakage (Samet et al., 18 Jul 2025).

The access model is only partially specified. The paper states “Dataset available here” and indicates that, if the URL is unavailable, access and further details can be requested from the organizers via the listed emails: Prof. Dr. Refik Samet, Zeynep Yildirim, Nooshin Nemati, Mohamed Traore, Assoc. Prof. Dr. Emrah Hancer, Prof. Dr. Serpil Sak, and Assoc. Prof. Dr. Bilge Ayca Kirmizi. License terms are not stated in the paper. The acknowledgement notes support by TUBITAK under Grant No. 121E379.

Several limitations are explicit. The sample size is 25 patients and 100 images, which the paper characterizes as sufficient for benchmarking but relatively modest compared to large-scale WSI datasets. Domain specificity is substantial: the dataset is exclusively breast NST. Acquisition variability is only partially characterizable because images originate from a single institution and pixel size is not reported. Annotation details are incomplete because the paper does not provide explicit guidelines, file format specification, or inter-annotator agreement statistics. Baseline results are absent, so community benchmarks must be established later.

These limitations do not negate the benchmark value of NuSeC; rather, they define the conditions under which its results should be interpreted. NuSeC is most rigorously understood as a compact, breast NST-specific nuclei segmentation dataset with manual QuPath annotations, explicit patient-level splitting, and an official evaluation protocol based on AJI and IoU.