---
title: BreakHis Dataset Overview
url: https://www.emergentmind.com/topics/breakhis-dataset
type: topic
---

# BreakHis Dataset Overview

BreaKHis, also written BreakHis in later papers, is a public breast histopathology benchmark used for automated analysis of microscopic biopsy imagery. In the literature surveyed here, it functions not only as a binary benign-versus-malignant classification dataset, but also as a substrate for 8-class subtype recognition, content-based medical image retrieval, case-based multi-magnification diagnosis, federated learning, augmentation studies, and uncertainty-aware cross-magnification evaluation [2305.03383]. Its enduring importance comes from a combination of multi-scale imaging, patient-structured data, and a strong malignant/benign imbalance, although the secondary literature is not fully standardized in how it reports image size, split protocol, or even some basic totals [2410.03333].

## 1. Identity and acquisition context

BreaKHis is consistently described as a breast tumor histopathology dataset composed of microscopic biopsy images, stained with hematoxylin and eosin or hematoxylin-eosin, and organized across four optical magnifications: \(40\times\), \(100\times\), \(200\times\), and \(400\times\) [1905.12005]. A detailed secondary description reports that the material was collected at the Pathological Anatomy and Cytopathology Laboratory in Paraná state, Brazil, during January to December 2014; that it includes both Surgical Open Biopsy and Core Needle Biopsy material; that sections are approximately \(3\,\mu\text{m}\) thick; and that imaging used an Olympus BX-50 microscope with a 3.3× magnification relay lens coupled to a Samsung SCC-131AN digital color camera [2209.01380].

The dataset is usually introduced as a benchmark for breast tumor diagnosis from microscopy images rather than as a whole-slide resource. Several papers explicitly treat the image as the native unit of analysis, while others conceptually frame the images as histopathology patches. This distinction matters because most BreakHis studies do not involve whole-slide reconstruction, slide-level region mining, or tile generation from WSIs; instead, they operate directly on the provided image set [2305.11728].

A central property of the dataset is that it contains many more images than patients. Multiple papers repeat that there are 82 patients in total. This creates a benchmark with nontrivial intra-patient correlation, which in turn makes split strategy a methodological issue rather than a minor implementation detail [1905.11567].

## 2. Composition and label structure

Across many later papers, the most stable dataset description is that BreaKHis contains **7,909** histopathology images from **82** patients, divided into **2,480 benign** and **5,429 malignant** images, with the following magnification-wise counts [2305.03383]:

| Magnification | Benign | Malignant / total |
|---|---:|---:|
| 40× | 625 | 1370 / 1995 |
| 100× | 644 | 1437 / 2081 |
| 200× | 623 | 1390 / 2013 |
| 400× | 588 | 1232 / 1820 |

The dataset is also repeatedly described as having an 8-subtype taxonomy: benign adenosis, fibroadenoma, phyllodes tumor, and tubular adenoma; malignant ductal carcinoma, lobular carcinoma, mucinous carcinoma, and papillary carcinoma [2603.14416]. One paper provides subtype totals as PC 560, MC 792, LC 626, DC 3451, TA 569, PT 453, F 1014, and A 444, again emphasizing the dominance of ductal carcinoma [2312.12804].

Later studies do not always use the full label structure. A large fraction of the literature collapses the dataset to a binary benign/malignant problem, even when the eight subtypes are acknowledged. Other works explicitly study the 8-class setting, and a smaller number define alternative tasks, including a 5-class setup consisting of one benign class plus four malignant carcinoma subclasses [2202.06109].

Secondary reporting is not fully consistent. Some papers report **2,368 benign** rather than 2,480 benign images, another reports **2,440 benign** images, and one paper reports **9,109 microscopic images** instead of 7,909 [1905.12005]. A separate study includes an obvious table error in which the 400× malignant count is printed as **5,429**, which is the dataset-wide malignant total rather than the 400× subset count [2209.01380]. This suggests that the dataset itself is widely used, but the surrounding literature has accumulated nontrivial transcription inconsistencies.

## 3. Imaging scale, representation, and reported image geometry

The four magnification levels are not merely metadata; they define much of the dataset’s scientific role. Some papers evaluate models independently at each of \(40\times\), \(100\times\), \(200\times\), and \(400\times\); some pool all magnifications into a single “combined” or “all-magnification” setting; and others deliberately restrict experiments to one scale, such as \(200\times\) or \(400\times\), to study a controlled regime [2410.03333]. This makes BreaKHis simultaneously a fixed-image benchmark and a scale-variation benchmark.

Reported native image size is another point of disagreement in the secondary literature. Several papers describe BreaKHis images as **700 × 460** RGB images [2309.01007]. Other papers report **752 × 582**, **720 × 360**, or even **224 × 224 × 3** as dataset image size [2403.11135]. In some cases these differences may reflect preprocessing or manuscript error rather than distinct dataset releases, but the later papers themselves do not fully reconcile them. A cautious reading is therefore that the most common nominal size in secondary reporting is \(700 \times 460\), while alternative resolutions also appear in the literature.

The magnification structure strongly affects methodological design. Compact texture-oriented CNNs have been tested at a single magnification, with the argument that different magnifications expose different balances of architecture and cellular detail [1905.12005]. Other studies explicitly target multi-resolution or cross-magnification robustness, treating BreaKHis as a benchmark for learning features that remain usable when magnification changes between training and deployment [2603.14416]. This suggests that BreaKHis has become, in practice, a benchmark for both classification and scale sensitivity.

## 4. Experimental protocols and benchmark regimes

No single evaluation protocol dominates the literature. Some studies follow a patient-wise protocol derived from earlier BreakHis practice, using repeated 70/30 train/test hold-outs with no patient overlap between training and testing [1905.11567]. Others use patient-wise 70/30 splits specifically to avoid leakage across multiple images from the same patient [2002.00072]. These designs are generally the most conservative because BreaKHis contains many correlated images per patient.

A large fraction of later work uses less clearly documented splits. Reported variants include 70/30 train/test, 60/20/20 train/validation/test, 80/20 train/test, and 85/15 train/test, often without explicit patient-level separation [2410.03333]. Several papers openly acknowledge the absence of a stated patient-wise split and note the resulting leakage risk, especially because training and test partitions may otherwise share images from the same patient [2504.12652]. Some papers also compare against prior methods evaluated under heterogeneous protocols, which limits fairness of benchmark claims.

Task formulation is equally heterogeneous. The dominant use is binary benign-versus-malignant classification. A second major use is subtype-level 8-class recognition. Beyond these, BreaKHis has been repurposed for top-\(K\) content-based medical image retrieval, where the test image is a query and retrieved gallery items are evaluated by label agreement rather than direct class prediction [2305.11728]. It has also been used for case-based diagnosis across all four magnifications, where one “case” contains one image from each magnification and yields a single patient-level malignancy decision [1905.11567].

Evaluation metrics follow the task. Classification papers commonly report accuracy and variants such as sensitivity, specificity, balanced accuracy, precision, recall, F1, and AUROC, although not all metrics are reported in every study. Retrieval papers instead use top-\(K\) precision, recall, F1, and related retrieval-oriented statistics, sometimes with query/gallery protocols that index training and validation images as prior cases while using every test image as a query [2305.03383].

## 5. Methodological roles in the literature

BreaKHis has served as a bridge between several methodological eras. Earlier work used handcrafted or hybrid pipelines based on decomposition, texture, and shape descriptors, such as iterative VMD with Zernike moments, fractal dimension, entropy features, ReliefF, and LS-SVM [1801.04880]. These studies treated the dataset as a structured texture-analysis problem and often emphasized interpretable descriptors over end-to-end learning.

The dataset then became a standard transfer-learning benchmark for CNNs and related architectures. Examples include IRRCNN for both binary and 8-class classification with image-level and patient-level reporting [1811.04241], comparative studies of VGG, ResNet, Inception-family, and Xception models on a magnification-specific subset [2309.01007], MobileNet-assisted lightweight classification pipelines [2403.11135], and later ViT-based transfer learning for per-magnification binary classification without data augmentation [2509.18553]. In this phase, BreaKHis was typically used to compare backbone families, fine-tuning depth, and deployment-oriented efficiency.

The dataset has also been repurposed in less conventional ways. One line of work uses it for content-based medical image retrieval rather than direct supervised classification, including unsupervised convolutional autoencoder retrieval, federated retrieval, and topological retrieval via persistent homology [2511.13170]. Another line uses it to study magnification-aware generalization, uncertainty, prototype reasoning, and mixed-magnification training, turning BreaKHis into a benchmark for scale shift and representation robustness rather than only lesion recognition [2603.14416].

Data augmentation research has used BreaKHis as a testbed for patient-aware balancing strategies. Gaussian-Laplacian pyramid blending was proposed specifically to combine stain appearance from different patients while balancing benign and malignant counts under patient-wise evaluation, reflecting the dataset’s simultaneous class imbalance and inter-patient variability [2002.00072]. This role is distinct from conventional architecture benchmarking: BreaKHis becomes a controlled environment for testing whether augmentation changes what a model learns.

## 6. Strengths, limitations, and benchmark interpretation

BreaKHis remains valuable because it combines several features rarely present together in a single small-to-medium histopathology benchmark: four magnification factors, benign/malignant labels, an 8-subtype taxonomy, and a patient-structured image collection. It supports binary classification, subtype recognition, retrieval, same-magnification and cross-magnification studies, and patient-level aggregation. This breadth explains why it appears in work ranging from handcrafted LS-SVM systems to federated CBMIR and transformer-based transfer learning [2305.03383].

Its main limitations are equally persistent. The patient count is small relative to the image count, creating a high risk of optimistic results under image-wise random splitting. Many later papers do not specify patient-wise partitioning, do not document validation protocol, or compare methods evaluated under incompatible split regimes [2403.11135]. Several studies also omit important implementation details such as optimizer settings, loss definitions, feature-layer selection, augmentation policy, or exact train/validation/test ratios, which weakens reproducibility [2504.12652].

A second limitation is the instability of secondary reporting. Conflicting statements about image counts, benign totals, nominal image size, and even per-magnification class counts appear across later papers. This suggests that the dataset’s scientific role is more stable than its secondary documentation. For benchmark use, the practical implication is that citation to “BreakHis performance” is meaningful only when paired with the exact task definition, magnification regime, split strategy, and evaluation metric [2410.03333].

The dataset’s continuing importance therefore lies less in a single canonical leaderboard than in its role as a common stress test for pathology methods under class imbalance, magnification variation, and patient-level dependence. A plausible implication is that BreaKHis is best understood not as one fixed experiment, but as a family of related benchmark regimes built on the same multi-magnification breast histopathology collection.

Source: https://www.emergentmind.com/topics/breakhis-dataset