SipakMed: Cervical Cytology Dataset

• SipakMed is a public dataset comprising 4,049 manually cropped Pap-smear images, annotated into five key cytomorphological classes.
• It has been used for various automated tasks including 5-class cell classification, segmentation, and nuclear-center localization with deep neural networks and vision transformers.
• The dataset highlights challenges in preprocessing and protocol consistency, such as variable crop sizes and differing annotation hierarchies.

SipakMed—also written SIPaKMeD or SiPaKMeD in the literature—is a public Pap-smear cervical cytology dataset used for automated analysis of cervical cells. Across the cited studies, it is most commonly described as containing 4,049 manually cropped single-cell images derived from a larger set of source images and annotated into five cytomorphological categories by expert cytopathologists. The dataset has been used for five-class cell classification, grouped abnormality classification, segmentation, nuclear-center localization, and risk-scoring experiments, and it appears in work spanning CNNs, hybrid deep-feature pipelines, ensembles, vision transformers, and multi-task architectures (Kundu et al., 2021, Liu et al., 2021).

1. Corpus, provenance, and naming conventions

Most studies describe SipakMed as a corpus of 4,049 isolated single-cell Pap-smear images together with 966 source images from which the cell crops were obtained. In the single-cell setting, the crops are already centered on individual cells, so many classification pipelines omit any explicit segmentation stage (Kundu et al., 2021, Khozaimi et al., 2024). The dataset is also described as publicly available on Kaggle in one optimizer-comparison study (Khozaimi et al., 2024).

The literature is not fully uniform in how it characterizes the source imagery. One study refers to 966 cluster-cell images and 4,049 manually cropped isolated-cell images (Kundu et al., 2021). A later segmentation-and-risk paper describes 966 high-resolution whole-slide images of size $2048\times1536$, from which 4,049 single-cell crops were manually extracted (Shaik et al., 21 May 2025). Another slide-analysis paper instead reports 204 non-overlapping fields of view of size $1280\times960$ acquired from Papanicolaou-stained slides, with the isolated-cell PNG crops exported from manually delineated ROIs (Panta et al., 29 Apr 2025). This suggests that downstream papers do not always preserve a single, fully consistent description of the acquisition hierarchy.

Original crop sizes are variable rather than standardized. One comparative deep-learning study reports a range from $62\times48$ to $531\times553$ pixels (Mustari et al., 2024), while CerviFormer reports a range from roughly $71\times59$ to $490\times474$ pixels (Deo et al., 2023). A robust-comparison paper emphasizes that this variability carries diagnostically relevant shape information, especially for cytomorphological cues such as the nucleus-to-cytoplasm ratio (Liu et al., 2021).

2. Cytomorphological taxonomy and annotation structure

The most frequently repeated five-class taxonomy in the cited literature is shown below.

Class	Count	Interpretation reported in the literature
Superficial-Intermediate	831	Mature squamous cells; normally predominant
Parabasal	787	Small immature basal-like cells
Koilocytotic	825	HPV-infected cells with perinuclear clearing
Dyskeratotic	813	Premature keratinization; associated with atypia
Metaplastic	793	Cells undergoing squamous metaplasia

This distribution is reported consistently in multiple studies using the single-cell subset (Liu et al., 2021, Basak et al., 2021, Kundu et al., 2021, Qian et al., 2023, Rahaman et al., 2021). The labels are morphological rather than purely diagnostic: two classes are commonly treated as normal, two as abnormal, and metaplastic as benign in grouped-class experiments. CerviFormer, for example, pools parabasal and superficial-intermediate into “Normal” (1,618 images), dyskeratotic and koilocytotic into “Abnormal” (1,638 images), and keeps metaplastic as “Benign” (793 images) for three-state classification (Deo et al., 2023).

Ground-truth labels are generally inherited from the released dataset rather than re-annotated by later authors. One transfer-learning study explicitly states that label assignment was inherited from the Kaggle release and that the authors did not perform additional relabeling, re-annotation, or expert-consensus procedures (Khozaimi et al., 2024). For segmentation-oriented work, additional annotation layers are reported. A Dense-UNet201 study states that ground-truth masks delineating nucleus and cytoplasm were available as $*.dat$ annotation files and were decoded into binary mask images (Khozaimi et al., 17 Apr 2025). A different multi-task segmentation study reports patch-level binary masks for source images, with white denoting cell and black denoting background (Shaik et al., 21 May 2025).

Not all later papers preserve the same class ontology. Two 2025 studies relabel the five categories as superficial squamous epithelial cells, intermediate squamous epithelial cells, columnar epithelial cells, low-grade squamous intraepithelial lesion cells, and high-grade squamous intraepithelial lesion cells, or very close variants of that scheme (Khozaimi et al., 17 Apr 2025, Khozaimi et al., 18 Jun 2025). Another 2024 study reports a different class breakdown, including Parabasal = 727 and Superficial-Intermediate = 813 (Mustari et al., 2024). This suggests that any cross-paper comparison should verify the exact label mapping used in the reported experiment.

3. Image geometry, preprocessing, and aspect-ratio handling

A central practical issue in SipakMed usage is normalization of highly variable cell-crop sizes to fixed network inputs. The most common target resolution is $224\times224$ pixels, used in transfer-learning and large-model comparison studies (Khozaimi et al., 2024, Liu et al., 2021). Other reported input sizes include $110\times110$ with aspect-ratio preservation (Mustari et al., 2024), $192\times192$ (Baba et al., 2024), $1280\times960$0 (Qian et al., 2023), $1280\times960$1 (Akash et al., 2024), and $1280\times960$2 (Panta et al., 29 Apr 2025).

Two broad preprocessing strategies recur. The first is direct resizing, which forces every crop to a fixed grid and distorts the original cell aspect ratio. The second is aspect-ratio-preserving square embedding, in which the shorter side is zero-padded before downsampling to the network input size (Liu et al., 2021). The systematic comparison in that study found that 12 of 22 models improved under direct resize and 10 decreased, with an average Scaled-minus-Standard accuracy change of $1280\times960$3, which the authors interpret as negligible (Liu et al., 2021). Their account attributes this robustness to the persistence of nuclear size, chromatin texture, and boundary detail even when whole-cell shape cues are distorted.

Normalization and augmentation practices vary substantially. Some studies report only resizing and generic “image augmentation,” without separate intensity normalization, color standardization, denoising, or filtering (Khozaimi et al., 2024). Others use ImageNet mean-standard-deviation normalization (Kundu et al., 2021), zero-mean unit-variance normalization (Rao, 5 Nov 2025), or domain-specific enhancement such as Perona-Malik diffusion and CLAHE (Khozaimi et al., 18 Jun 2025). Augmentation ranges from none explicitly reported (Kundu et al., 2021) to extensive affine, photometric, noise-injection, zoom, padding, and contrast perturbations (Mustari et al., 2024, Rahaman et al., 2021).

Because the single-cell subset is already pre-cropped, many classification studies treat SipakMed as a direct recognition problem rather than a detect-then-classify problem (Khozaimi et al., 2024). By contrast, localization work derives a nuclear centroid from contour annotations and then pads each image to a $1280\times960$4 frame for coordinate regression (Rao, 5 Nov 2025).

4. Experimental protocols and task formulations

Evaluation on SipakMed is protocol-dependent. Reported splits include $1280\times960$5 train/validation/test (Liu et al., 2021, Kundu et al., 2021), $1280\times960$6 (Khozaimi et al., 2024, Khozaimi et al., 18 Jun 2025), $1280\times960$7 train/test without a distinct validation set (Baba et al., 2024), $1280\times960$8 on an augmented dataset (Akash et al., 2024), and five-fold cross-validation (Ahmed, 15 Aug 2025, Panta et al., 29 Apr 2025, Rao, 5 Nov 2025). Reported metrics therefore are not directly commensurate across studies.

The dataset supports several task formulations.

Task	Representative formulation	Representative metrics
Single-cell classification	5-class, 3-class, or 2-class	Accuracy, precision, recall, $1280\times960$9, loss, AUC
Segmentation	Pixel-wise nucleus/cytoplasm or cell/background masks	IoU, Dice, accuracy, loss
Localization	Nuclear-center regression from contour centroid	$62\times48$0, MSE
Risk scoring	Class-conditional likelihoods from learned features	Posterior probabilities, cosine similarity

Five-class classification is the dominant use case, but grouped-label settings are common. DeepCervix reports 2-class, 3-class, and 5-class experiments on the same base dataset (Rahaman et al., 2021). CerviFormer uses the three-state Normal/Abnormal/Benign grouping (Deo et al., 2023). Whole-slide or cluster-image classification also appears in the Sugeno-fusion study, which treats both the 966-image WSI subset and the 4,049-image SCI subset as separate benchmarks (Kundu et al., 2021).

Segmentation formulations are less standardized. One paper uses resized $62\times48$1 RGB images and corresponding binary JPG masks decoded from $62\times48$2 files (Khozaimi et al., 17 Apr 2025). Another uses $62\times48$3 overlapping patches extracted from source images, discarding all-black patches and learning a multi-task U-Net for simultaneous segmentation and classification (Shaik et al., 21 May 2025). A localization study instead treats each sample as an image-annotation pair and regresses the centroid $62\times48$4 of the annotated nuclear contour (Rao, 5 Nov 2025).

5. Benchmark results and methodological findings

Reported performance on SipakMed spans a wide range, and the range itself is informative because it reflects changes in split protocol, label grouping, preprocessing, augmentation, and model family. Under an $62\times48$5 protocol with ResNet-18, ResNet-34, and VGG-16, one transfer-learning study reports best accuracies of $62\times48$6 for VGG-16 with Adamax, $62\times48$7 for ResNet-18 with Adamax, and $62\times48$8 for ResNet-34, with Nadam exceeding Adamax on ResNet-34 by $62\times48$9 (Khozaimi et al., 2024). In the same study, the central conclusion is that transfer learning improves all CNNs and optimizers and that optimizer choice has little influence once transfer learning is used (Khozaimi et al., 2024).

On a $531\times553$0 split across 22 deep models, the aspect-ratio comparison study reports that GoogLeNet achieved $531\times553$1 accuracy on directly resized data, DeiT $531\times553$2, DenseNet169 $531\times553$3 on aspect-ratio-preserving standard data, and DeiT $531\times553$4 on standard data (Liu et al., 2021). Cost-sensitive and supervised-contrastive training on the single-cell subset yields $531\times553$5 accuracy, with the paper reporting that class weighting modestly improved minority-class recall and that contrastive pre-training improved inter-class separability, especially between dyskeratotic and koilocytotic cells (Mustari et al., 2024). A hybrid pipeline based on deep feature extraction, PCA preserving $531\times553$6 variance, Grey Wolf Optimizer feature selection, and RBF-SVM reports $531\times553$7 test accuracy (Basak et al., 2021).

Ensemble methods push reported accuracy higher under their own protocols. A Sugeno fuzzy-integral fusion of Inception v3, DenseNet-161, and ResNet-34 reaches $531\times553$8 accuracy on the SCI subset and $531\times553$9 on the WSI subset (Kundu et al., 2021). A Voting-Stacking ensemble of three Inception-family models reports $71\times59$0, $71\times59$1, and $71\times59$2 accuracy under $71\times59$3, $71\times59$4, and $71\times59$5 test hold-outs, with a three-level super-ensemble reaching $71\times59$6 in the $71\times59$7-hold-out setting (Qian et al., 2023). DeepCervix, which fuses $71\times59$8-dimensional pre-softmax features from VGG16, VGG19, ResNet50, and Xception into a $71\times59$9-dimensional representation, reports $490\times474$0 for 2-class, $490\times474$1 for 3-class, and $490\times474$2 for 5-class classification (Rahaman et al., 2021).

Transformer-based approaches show similarly heterogeneous but competitive behavior. CerviFormer applies cross-attention and a latent-transformer bottleneck to $490\times474$3 images partitioned into $490\times474$4 patches and reports $490\times474$5 accuracy for 3-state classification, along with $490\times474$6 sensitivity, $490\times474$7 specificity, and Cohen’s $490\times474$8 (Deo et al., 2023). HistoViT fine-tunes a ViT-Base patch16 model and reports $490\times474$9 accuracy, $*.dat$0 mean precision, $*.dat$1 mean recall, and $*.dat$2 AUC under five-fold cross-validation (Ahmed, 15 Aug 2025). A CvT-13 classifier trained for low-resource slide analysis reports $*.dat$3 accuracy and average AUC $*.dat$4 under five-fold cross-validation when fine-tuned from ImageNet-22k (Panta et al., 29 Apr 2025).

SipakMed has also supported non-classification tasks. A multi-task U-Net study reports IoU $*.dat$5 for segmentation and $*.dat$6 accuracy for simultaneous classification, while its standalone Multi-Resolution Fusion Deep Convolutional Network reaches $*.dat$7 classification accuracy using about $*.dat$8 million parameters (Shaik et al., 21 May 2025). A Dense-UNet201 plus Spider Monkey Optimization study reports segmentation accuracy $*.dat$9, IoU $224\times224$0, and Dice $224\times224$1 (Khozaimi et al., 17 Apr 2025). For nuclear-center localization, replacing pure CoordConv channels with intensity-weighted coordinate channels improves mean test-set $224\times224$2 by $224\times224$3 with a $224\times224$4 bootstrap CI of $224\times224$5 and one-sided $224\times224$6 (Rao, 5 Nov 2025).

6. Limitations, inconsistencies, and research significance

Several constraints recur across SipakMed-based studies. The total single-cell sample count of 4,049 is small for end-to-end deep learning, which is why transfer learning, augmentation, and feature reuse appear so frequently (Khozaimi et al., 2024). The dataset is fundamentally a single-cell crop collection; one study explicitly notes that no clinical metadata or multi-cell fields of view are integrated and that the dataset is “strictly single-cell crops” (Khozaimi et al., 2024). A plausible implication is that SipakMed is strongest as a controlled cytomorphology benchmark rather than a complete proxy for slide-level screening workflows.

The literature also exposes several interpretive pitfalls. Source-image descriptions differ, class ontologies are not always preserved, and train/validation/test protocols vary substantially (Shaik et al., 21 May 2025, Khozaimi et al., 18 Jun 2025). Some papers omit details on microscope magnification, staining variants, stratification, validation protocol, or class-imbalance handling (Baba et al., 2024, Khozaimi et al., 2024). This makes raw leaderboard-style comparison unreliable unless preprocessing, grouping, augmentation, and evaluation schemes are aligned.

At the same time, the dataset has been unusually productive methodologically. It has been used to test whether aspect-ratio preservation matters (Liu et al., 2021), whether cost-sensitive and contrastive learning improve minority recall and representation geometry (Mustari et al., 2024), whether hybrid deep features and swarm-based optimization outperform direct classifiers (Basak et al., 2021, Khozaimi et al., 17 Apr 2025), and whether lightweight models can approximate larger backbones at lower computational cost (Shaik et al., 21 May 2025, Akash et al., 2024). The breadth of these uses indicates that SipakMed functions not only as a classification corpus but also as a compact experimental platform for studying preprocessing, optimization, explainability, segmentation, localization, and low-resource deployment in cervical cytology.