MiDeSeC Dataset for Mitosis Detection & Segmentation

• MiDeSeC is a publicly available histopathology image dataset containing H&E stained invasive breast carcinoma images for mitosis detection and segmentation.
• It features 50 high-power fields from 25 patients with over 500 mitoses annotated at pixel level, supporting both detection and segmentation tasks.
• The dataset addresses challenges in breast cancer grading by enabling the development of automated methods that reduce subjectivity and improve reproducibility.

Searching arXiv for the MiDeSeC paper and closely related mitosis-detection context papers. arXiv search: MiDeSeC (Samet et al., 18 Jul 2025) and related mitosis benchmark datasets. Searching arXiv for ([2507.14271](/papers/2507.14271)). [arXiv search] query: (Samet et al., 18 Jul 2025) MiDeSeC is a publicly available histopathology image dataset for mitosis detection and mitosis segmentation in breast cancer, introduced in “MiDeSeC: A Dataset for Mitosis Detection and Segmentation in Breast Cancer Histopathology Images” (Samet et al., 18 Jul 2025). It consists of H&E stained invasive breast carcinoma, no special type (NST) material from 25 different patients, captured at $40\times$ magnification, with 50 regions of interest of size $1024 \times 1024$ pixels and more than 500 mitoses in total. The dataset is designed around the observation that mitotic count is a central component of breast cancer grading and that mitotic figures are rare, morphologically diverse, and difficult to identify reproducibly in routine manual review.

1. Clinical role and problem formulation

MiDeSeC is explicitly intended for two tasks: mitosis detection, defined as finding where mitotic figures are located, and mitosis segmentation, defined as delineating the pixel-level extent or shape of each mitotic figure (Samet et al., 18 Jul 2025). These tasks are positioned as central to breast cancer grading. In the Nottingham Grading System for breast cancer, the three key morphological criteria are mitotic count, tubule formation, and nuclear pleomorphism, and mitotic count is described as the most prominent feature.

The dataset is motivated by several practical and methodological difficulties in histopathology. Pathologists routinely scan Hematoxylin and Eosin stained slides at high-power fields to count mitoses, but there are many high-power fields per slide, making manual screening tedious and time-consuming. Mitoses are rare relative to non-mitotic nuclei, and they occur in multiple stages—prophase, metaphase, anaphase, and telophase—with substantially different shapes and structures. In addition, some non-mitotic cells, including apoptotic cells and dense nuclei, can closely resemble mitoses, introducing subjectivity and inter-observer variability.

Within this framing, MiDeSeC is intended to support automatic algorithms that assist or partially automate mitotic count, contribute to more robust and reproducible cancer grading systems, and reduce workload and variability in routine pathology. A plausible implication is that the dataset is best understood not merely as an image collection but as a task-oriented resource for quantitative computational pathology.

2. Acquisition protocol and dataset composition

The samples in MiDeSeC are derived from invasive breast carcinoma, no special type (NST), from the Department of Medical Pathology, Ankara University, Ankara, Turkey (Samet et al., 18 Jul 2025). All slides are H&E-stained. The text does not specify any special stain normalization procedure or any variant of the H&E protocol; it implies standard clinical H&E staining.

Slides were scanned by a 3D Histech Panoramic p250 Flash-3 scanner and an Olympus BX50 microscope at $40\times$ magnification. The exact pixel size in micrometers per pixel is not specified. The text notes, however, that given the $40\times$ magnification, each region is a high-power field suitable for mitosis assessment.

The overall dataset structure is compact and tightly defined:

Component	Value	Note
Patients	25	Invasive breast carcinoma, NST
Regions of interest / HPFs	50	Selected from glass slides
ROI size	$1024 \times 1024$ pixels	Fixed-size fields
Total mitoses	More than 500	Across the 50 regions
Split	Two-thirds training, one-third testing	Region-level split

The regions were selected to capture “several possible mitosis shapes,” reflecting different morphological appearances and stages of mitosis. The exact sampling criteria, including the balance among tumor, background, and stroma, are not detailed. The paper states only that the focus is on areas with relevant tumor tissue and mitotic activity. Because there are more than 500 mitoses in 50 regions, a rough average exceeds 10 mitoses per image, although no distributional statistics such as mean, median, or variance are reported.

The train–test partition is performed at the region level rather than the patch level. The paper does not explicitly confirm patient-level exclusivity in the split. Since 50 regions are drawn from 25 patients, it is likely that some patients contribute more than one region. This suggests that studies requiring strict subject-level separation may need to define an additional patient-level protocol.

3. Annotation schema and ground-truth representation

Each ROI is accompanied by a CSV text file that constitutes the ground truth (Samet et al., 18 Jul 2025). The CSV has no header line. Each line corresponds to one mitosis, and each line provides the coordinates of all pixels belonging to that mitosis. The coordinate system is defined with origin at the top-left corner, $(0,0)$, increasing rightwards in $x$ and downwards in $y$. Pixel coordinates are integers, and any floating-point coordinates are rounded; the example given states that a pixel specified as $(7.89, 3.02)$ is displayed at $(7, 3)$.

The paper’s figure illustrates the annotation format with explicit coordinate lists. One example mitosis is represented by the blue pixels

$1024 \times 1024$0

with center $1024 \times 1024$1. Another is represented by the red pixels

$1024 \times 1024$2

with center $1024 \times 1024$3.

A notable property of the annotation model is that a single mitosis may have a disconnected mask. The paper states that while the red structure represents a single mitosis, the two blue structures also represent a single mitosis; in other words, some mitoses may involve a gap in their shapes. This is technically important because it means the unit of annotation is the mitotic instance rather than a single connected component.

The annotation design directly supports reconstruction of pixel-level segmentation masks and derivation of instance masks, one mask per mitosis. It also supports center-point extraction for detection tasks, for example by computing the centroid of the listed pixel coordinates. The paper involves expert pathologists as co-authors from the Department of Medical Pathology, Ankara University, but it does not specify how many pathologists annotated the data, whether consensus rounds or adjudication were used, or whether quantitative inter-observer measures such as Cohen’s kappa were computed.

4. Detection and segmentation tasks and evaluation definitions

Because each mitosis is annotated at pixel level, MiDeSeC directly supports both segmentation and detection (Samet et al., 18 Jul 2025). For segmentation, the task is to predict the set of pixels belonging to each mitosis. For detection, mask centers or connected components can be used to define mitosis locations.

The paper defines the evaluation in terms of true positives, false positives, and false negatives. Specifically, $1024 \times 1024$4 denotes the number of true positives, that is, detected mitoses that correspond to ground truth mitoses; $1024 \times 1024$5 denotes false positives, that is, detections that are not ground truth mitoses; and $1024 \times 1024$6 denotes false negatives, that is, ground truth mitoses not detected. The reported formulas include

$1024 \times 1024$7

and

$1024 \times 1024$8

The paper also defines an $1024 \times 1024$9-Score in terms of precision and recall.

At the same time, the official matching rule is not specified. The text does not define a distance threshold between centers, an intersection-over-union threshold between masks, or any other criterion that determines when a prediction counts as a true positive. It also does not provide official pixel-wise segmentation metrics such as Dice or IoU, although it notes that these may be appropriate to use for segmentation tasks. This absence of a formal matching protocol is one of the dataset’s main methodological open points: the basic count-based metrics are explicit, but the operationalization of instance correspondence is left to downstream studies.

The paper further does not provide baseline algorithms, benchmark architectures, performance tables, or numerical results. No CNN, U-Net, Mask R-CNN, or classical reference method is reported. A common misconception would therefore be to treat MiDeSeC as an established leaderboard benchmark with canonical baseline scores. The source paper presents it primarily as a data resource with annotation and evaluation definitions, rather than as a completed benchmark suite.

5. Relation to prior mitosis datasets and benchmark culture

MiDeSeC is introduced in the context of several earlier mitosis detection challenges: the ICPR 2012 mitosis detection contest, AMIDA13 at MICCAI 2013, MITOS-ATYPIA at ICPR 2014, and TUPAC16 at MICCAI 2016 (Samet et al., 18 Jul 2025). The paper states that many previous contests do not specifically focus on breast cancer, except TUPAC16.

Within that landscape, MiDeSeC is characterized as being specifically designed for breast cancer histopathology and for the dual tasks of detection and segmentation in H&E images. The paper does not present a detailed comparative table covering number of cases, magnification, exact resolution, annotation type, or scanner and stain variability. It nonetheless identifies several distinguishing features from the described content: explicit pixel-level segmentation masks for mitoses, dedication to invasive breast carcinoma NST under H&E at $40\times$0, and consistent acquisition conditions.

This suggests that MiDeSeC occupies a more specialized niche than some older mitosis challenges. Its annotation granularity is particularly relevant for instance-level segmentation pipelines, whereas many earlier mitosis datasets are commonly associated with point-level supervision. Because the paper does not provide a systematic side-by-side comparison, this characterization should be read as a description of design intent rather than as a fully audited benchmark taxonomy.

6. Access, practical use, and limitations

The paper states that the dataset can be downloaded from an online link, phrased as “The MiDeSeC dataset can be downloaded from the following link: Dataset available here” (Samet et al., 18 Jul 2025). The extracted text does not include the URL itself. Distribution is supported by the organizing institutions, including Ankara University and Mehmet Akif Ersoy University, and funded by TUBITAK grant 121E379. The paper does not specify the exact license, and it does not provide explicit IRB or ethical approval details in the extracted text.

For practical use, the source text outlines a workflow centered on loading each $40\times$1 color ROI together with its CSV annotation, reconstructing binary or instance masks from the pixel coordinate lists, and optionally extracting center points for detection. It further notes that no preprocessing or augmentation protocol is mandated by the dataset. There is no mention of color normalization, stain normalization, artifact removal, or patch tiling algorithms beyond ROI selection. The fixed-size high-power fields make additional tiling unnecessary unless a model requires smaller inputs.

Several limitations are either explicit or implicit in the source description. MiDeSeC is single-center data from Ankara University, which limits variability in staining and scanning protocols. It covers a single disease subtype— invasive breast carcinoma, NST only—rather than multiple histologic subtypes or multiple cancers. It includes 25 patients and 50 HPFs, which is modest relative to large-scale whole-slide imaging resources. Documentation of scanner or staining variability is limited, despite mention of two acquisition devices. There are no official baseline models or scores, and no detailed annotation-process audit is reported.

These constraints have direct methodological consequences. Single-center acquisition can introduce domain bias; absence of patient-level split guarantees can complicate strict generalization studies; and lack of official matching criteria means that different groups may report non-comparable instance-level scores. Conversely, the combination of breast-cancer specificity, high-power-field imaging, and per-instance pixel coordinate annotations makes MiDeSeC a compact but technically rich resource for developing and testing mitosis detection and segmentation methods in breast cancer histopathology.