ICIAR2018 Breast Cancer Histology Benchmark

• ICIAR2018 is a breast histology benchmark that categorizes H&E-stained images into normal, benign, in situ carcinoma, and invasive carcinoma.
• It supports diverse methodological approaches including CNNs, transformers, and graph-based models to tackle small-data challenges and emphasize interpretability.
• Reported results vary with protocols, showcasing accuracies from around 87% to near-perfect performance depending on preprocessing and evaluation strategies.

ICIAR2018 commonly denotes the BACH breast cancer histology benchmark from the ICIAR 2018 Grand Challenge on Breast Cancer Histology Images, a publicly available histopathological dataset used for four-class breast tissue classification into normal, benign, in situ carcinoma, and invasive carcinoma. In the literature, it functions both as an original microscopy-image benchmark and, in later work, as a tiled corpus derived from larger histology sources. This dual use has made ICIAR2018 a recurring testbed for small-data generalization, patch-based and image-level representation learning, transformer-based modeling, and interpretable graph reasoning in computational pathology (Rakhlin et al., 2018, Nazeri et al., 2018, Ahmed, 15 Aug 2025, Gao et al., 10 Feb 2026).

1. Benchmark identity and diagnostic categories

In the challenge-oriented literature, ICIAR2018 is defined as a breast histopathology classification task over four tissue categories: normal, benign, in situ carcinoma, and invasive carcinoma. Multiple papers describe the benchmark as containing 400 H&E-stained images, with balanced class membership and one label per image corresponding to the predominant cancer type. One study specifies that the images are $2048 \times 1536$ pixels, acquired under the same imaging conditions at $200\times$ magnification with pixel size $0.42\ \mu m \times 0.42\ \mu m$, and annotated by two medical experts (Rakhlin et al., 2018). Another graph-based study restates the four-class structure as 100 images per class, emphasizing the benchmark’s balanced design and small size (Gao et al., 10 Feb 2026).

The task is technically nontrivial because breast histology images are both large and morphologically heterogeneous. The two-stage CNN study characterizes the problem as requiring both local cell-level detail and global tissue structure, while the graph-based literature describes the categories as differing by fine-grained tissue morphology rather than coarse global appearance (Nazeri et al., 2018, Gao et al., 10 Feb 2026). This combination of balanced labels, limited sample count, and subtle class boundaries has made ICIAR2018 a canonical benchmark for evaluating whether a method can avoid overfitting while preserving diagnostically relevant structure.

2. Dataset realizations and task formulations

The dominant use of ICIAR2018 is the four-class benchmark described above. A secondary formulation collapses the labels into a two-class carcinoma-detection task, with non-carcinoma defined as normal plus benign and carcinoma defined as in situ plus invasive carcinoma (Rakhlin et al., 2018). This binary variant is relevant when the diagnostic objective is screening for malignancy rather than subtype discrimination.

A source of terminological ambiguity is that later work operationalizes ICIAR2018 in different input representations. Earlier challenge papers work directly from the original microscopy images, whereas the HistoViT study describes ICIAR2018 as the BACH breast cancer histology benchmark and uses 11,794 image tiles extracted from 400 H&E-stained whole-slide images by an overlapping sliding window approach, with each tile sized $512 \times 512$ and required to contain at least 80% tissue content based on color pixel thresholds (Ahmed, 15 Aug 2025). The same benchmark name therefore refers either to the original image-level dataset or to a tile-level derivative used for deep learning experiments.

Preprocessing conventions likewise vary substantially across studies. One CNN-plus-boosting pipeline applies Macenko stain normalization, 50 random color augmentations using Ruifrok color deconvolution, and multi-scale cropping from both $2048 \times 1536$ and $1024 \times 768$ images (Rakhlin et al., 2018). By contrast, HistoViT uses a lightweight pipeline in which RGB images are normalized to $[0,1]$, converted to PyTorch’s channel-first tensor format, resized to $224 \times 224$, and stacked into batches of size 32, with no data augmentation (Ahmed, 15 Aug 2025). GAFR-Net also resizes images to $224 \times 224$ and normalizes them, but adds horizontal flip, vertical flip, and random rotations within $\pm 30^\circ$ (Gao et al., 10 Feb 2026). These differences are central to interpreting reported performance.

3. Methodological trajectories on ICIAR2018

A prominent early line of work treats ICIAR2018 as a hierarchical visual recognition problem. The two-stage CNN model divides an image $200\times$0 into patches,

$200\times$1

learns patch descriptors through a patch-wise encoder,

$200\times$2

and then aggregates these patch features into an image-level prediction,

$200\times$3

In that formulation, the first network acts as an autoencoder that extracts salient local features, while the second image-wise CNN captures global organization across patches (Nazeri et al., 2018). The architectural premise is that histopathology diagnosis depends on both microscopic structures and their tissue-scale arrangement.

A second trajectory replaces end-to-end classification with feature extraction plus tabular learning. In that approach, ImageNet-pretrained ResNet-50, InceptionV3, and VGG-16 act as fixed feature encoders, and LightGBM performs final classification on pooled image descriptors (Rakhlin et al., 2018). The pipeline uses 3-norm pooling over crop descriptors, extensive augmentation, and large-scale model averaging. The authors explicitly report that fine-tuning large pretrained CNNs did not perform well on the small dataset, which motivated the use of pretrained deep features and gradient boosted trees instead.

Transformer-based work reframes the benchmark around global-context modeling. HistoViT instantiates ViT-Base (vit-base-patch16-224) from Hugging Face, pretrained on ImageNet and then fine-tuned for $200\times$4 classes (Ahmed, 15 Aug 2025). Its tokenization follows the standard patch-projection scheme

$200\times$5

after which positional encodings and a learnable classification token are prepended before Transformer processing. Training minimizes cross-entropy with Adam. In the authors’ interpretation, self-attention is especially effective at modeling long-range spatial dependencies and subtle tissue morphology, which are central in breast histopathology.

An interpretable alternative appears in graph-based modeling. GAFR-Net constructs a similarity graph $200\times$6 over images, adding edges when

$200\times$7

For each node $200\times$8, it computes a topology vector

$200\times$9

where $0.42\ \mu m \times 0.42\ \mu m$0 is clustering coefficient, $0.42\ \mu m \times 0.42\ \mu m$1 is node degree, and $0.42\ \mu m \times 0.42\ \mu m$2 is two-hop label agreement. Message passing uses multi-head graph attention,

$0.42\ \mu m \times 0.42\ \mu m$3

and a differentiable fuzzy-rule module computes rule activations through

$0.42\ \mu m \times 0.42\ \mu m$4

The fused representation is

$0.42\ \mu m \times 0.42\ \mu m$5

followed by softmax classification (Gao et al., 10 Feb 2026). This design integrates relational learning with explicit symbolic reasoning.

4. Protocol design and sources of non-comparability

Reported ICIAR2018 results are strongly protocol-dependent. The CNN-plus-boosting study uses 10 stratified folds, keeps all descriptors from the same original image in the same fold to prevent leakage, repeats LightGBM training with 5 random seeds, and averages predictions across models and augmentations (Rakhlin et al., 2018). The two-stage CNN paper reports performance on the validation set, but the provided text does not expose exact patch size, stride, augmentation policy, optimizer, learning rate, epoch counts, or split ratios (Nazeri et al., 2018).

HistoViT adopts an 80:20 train:test split chosen to match prior work and ensure fair comparison, trains for up to 50 epochs with Adam at learning rate $0.42\ \mu m \times 0.42\ \mu m$6, batch size 32, cross-entropy loss, and early stopping based on test accuracy with patience $0.42\ \mu m \times 0.42\ \mu m$7, and explicitly uses no data augmentation (Ahmed, 15 Aug 2025). GAFR-Net, by contrast, trains in PyTorch 2.0.1 with learning rate $0.42\ \mu m \times 0.42\ \mu m$8, batch size 16, 50 epochs, and repeats each experiment four times, but the provided text does not explicitly specify train/validation/test split ratios for ICIAR2018 (Gao et al., 10 Feb 2026).

A common misconception is that ICIAR2018 numbers can be read as if they were generated under a single evaluation regime. The published protocols summarized above do not support that assumption. This suggests that raw score comparisons should be interpreted alongside split design, augmentation policy, tile-versus-image representation, and whether evaluation is image-level, validation-only, or cross-validated.

5. Reported results in representative studies

The benchmark has supported a wide range of reported results, reflecting both architectural progress and evaluation heterogeneity.

Study	Method and protocol	Reported ICIAR2018 result
Two-stage CNN (Nazeri et al., 2018)	Patch-wise autoencoder plus image-wise CNN; validation-set evaluation	95% accuracy
CNN features + LightGBM (Rakhlin et al., 2018)	ImageNet-pretrained ResNet-50, InceptionV3, VGG-16; 10-fold stratified CV	87.2% ± 2.6 accuracy for 4-class; 93.8% ± 2.3 accuracy and 0.973 AUC for 2-class
HistoViT (Ahmed, 15 Aug 2025)	ViT-Base, ImageNet pretrained, 80:20 train:test, no augmentation	Accuracy 99.32%; Precision 99.32%; Recall 99.32%; AUC 99.95
GAFR-Net (Gao et al., 10 Feb 2026)	Graph attention plus fuzzy rules; experiments repeated 4 times	AUC-ROC $0.42\ \mu m \times 0.42\ \mu m$9; Balanced Accuracy $512 \times 512$0; F1 $512 \times 512$1; Kappa $512 \times 512$2; Sensitivity 80.24%; Specificity 81.36%

The HistoViT comparison table for breast cancer situates its 99.32% result above a set of earlier methods reported at 79.00% for Kwok et al. with a 75:25 split, 81.25% for Nawaz et al. with 80:20, 87.20% for Rakhlin et al. with 80:20, 87.50% for Vang et al. with 75:25, 92.50% for DCNN with 75:25, and 91.64% for TDA with 80:20. The same paper reports a confusion matrix for breast cancer and states qualitatively that the model captures the “complex morphological distinctions among benign, in-situ, and invasive tumor classes,” while the AUC figure is used to argue exceptional one-vs-rest discriminability, though per-class AUC values are not listed in the text.

The GAFR-Net comparison table reports higher AUC, balanced accuracy, F1, and Kappa than the baselines listed there, including ResNet50, GCN, ViT, Swin-Tiny, CoAtNet, GAT, PFE-INC-RES, HoRFNet, and CellSage. Its ablation on ICIAR2018 further reports performance drops when graph attention, the attention mechanism, or fuzzy rules are removed, indicating that the full graph-attention-plus-fuzzy-rule combination is the strongest configuration in that study.

6. Interpretability, robustness, and continuing significance

ICIAR2018 has persisted as more than a simple four-class benchmark because it compresses several core problems of computational pathology into a small, controlled setting. The two-stage CNN work frames the task around local evidence and global evidence, the boosted-tree pipeline emphasizes overfitting risk on very small datasets, the transformer study argues for long-range dependency modeling, and the graph-based approach treats inter-sample relationships and explanation as first-class design objectives (Nazeri et al., 2018, Rakhlin et al., 2018, Ahmed, 15 Aug 2025, Gao et al., 10 Feb 2026).

Interpretability is especially explicit in GAFR-Net, whose fuzzy-rule module encodes graph topology into human-readable logic. The representative rule reported in the paper is: IF node degree is HIGH AND label agreement is STRONG, THEN the sample belongs to a HIGH-CONFIDENCE region (Gao et al., 10 Feb 2026). This differs from post-hoc attribution, since the explanation is embedded in the forward pass through trainable Gaussian membership functions over node degree, clustering coefficient, and two-hop label agreement.

Robustness claims are likewise benchmark-specific. HistoViT emphasizes that ICIAR2018 is relatively small and that no data augmentation is used, arguing that strong performance under these conditions demonstrates robustness and reliability of the learned representation (Ahmed, 15 Aug 2025). Because that study uses only normalization, resizing, patchification, and fine-tuning of an ImageNet-pretrained ViT-Base model, it positions ICIAR2018 as a particularly clear demonstration of global-attention efficacy for multi-class breast cancer histology classification.

Taken together, these studies suggest that ICIAR2018 functions as a compact stress test for several competing inductive biases: local patch encoders, pooled pretrained CNN descriptors, global self-attention, and graph-topological reasoning. Its continued use reflects not only the diagnostic relevance of its four tissue categories, but also its suitability for probing how learning systems behave under limited data, subtle morphology, and clinically motivated demands for explanation.