SICAPv2 Prostate Histopathology Benchmark

• SICAPv2 is a comprehensive prostate histopathology benchmark built from H&E-stained biopsies with detailed Gleason-grade annotations for varied computational tasks.
• The dataset supports multiple task-specific pipelines, including patch classification, slide-level grading, and cribriform pattern detection, with diverse preprocessing protocols.
• Its flexible annotation schema and safety-aware evaluations highlight methodological trade-offs and standardization challenges in computational prostate pathology.

SICAPv2 is a prostate histopathology benchmark built around H&E-stained biopsy whole-slide images and Gleason-grade annotations. In its foundational description, it extends the earlier SICAP resource and couples whole-slide Gleason scoring with local Gleason pattern annotations and explicit cribriform markings, thereby supporting patch-level grading, cribriform detection, and biopsy-level score prediction within a single corpus (Silva-Rodríguez et al., 2021). Subsequent work has reused SICAPv2 under multiple task formulations—including four-class patch classification, binary benign–malignant safety analysis, five-class weakly supervised ISUP grading, CBMIR retrieval, and multi-class segmentation—so the name denotes both a source dataset and a family of derived experimental instantiations in computational prostate pathology (Mohammadi-Seif et al., 4 May 2026, Mohammadi-Seif et al., 14 Apr 2026, Zaffar et al., 2022).

1. Foundational dataset and reported corpus structure

The original SICAPv2 description reports 155 biopsies from 95 patients, digitized as WSIs, with a total of 182 WSIs because some patients contribute more than one biopsy (Silva-Rodríguez et al., 2021). At the WSI level, the same source gives the following counts by primary Gleason grade: NC: 37, GG3: 60, GG4: 69, and GG5: 16, for 182 WSIs in total (Silva-Rodríguez et al., 2021). The images are H&E stained prostate needle biopsies scanned with a Ventana iScan Coreo at $40\times$ and then down-sampled to $10\times$ for computational processing (Silva-Rodríguez et al., 2021).

In the original end-to-end grading pipeline, the downsampled WSIs are tiled into $512 \times 512$ patches with 50\% overlap (Silva-Rodríguez et al., 2021). After preprocessing and selection, the patch-level composition is reported as 10,340 patches: 4,417 NC, 1,636 GG3, 3,622 GG4, and 665 GG5; within GG4, 763 patches are annotated as cribriform (Silva-Rodríguez et al., 2021). A later classification study using SICAPv2 reports the same four-class patch count of 10,340, with 4417 benign, 1636 grade 3, 3622 grade 4, and 665 grade 5 patches, extracted as $512 \times 512$ tiles with 50\% overlap from biopsy WSIs (Malekmohammadi et al., 2024).

Other studies materialize SICAPv2 differently. In an attention-based MIL setting for WSI-level ISUP grading, SICAPv2 is described as 155 whole-slide images of prostate biopsies, processed at $10\times$ magnification, with an average of 129 patches per WSI and $256 \times 256$ patch size (Zaffar et al., 2022). A ViT-based classification study instead reports 11,836 image patches distributed across NC: 2,046, G3: 2,613, G4: 5,768, and G5: 1,409, using an 80:20 split (Ahmed, 15 Aug 2025). A CBMIR study gives 10,330 patches for retrieval experiments and uses 2,122 test patches as queries (Tabatabaei et al., 2023). These parallel descriptions indicate that SICAPv2 is not used in a single canonical tensorized form; rather, studies derive task-specific sample sets from the same prostate biopsy collection.

2. Annotation schema and task-dependent label spaces

In the original dataset design, SICAPv2 combines biopsy-level Gleason scores, pixel-level Gleason grade annotations, and explicit cribriform pattern markings within Gleason grade 4 regions (Silva-Rodríguez et al., 2021). Patch-level labels are obtained from local annotations by assigning the majority grade when more than one annotated grade is present in a patch (Silva-Rodríguez et al., 2021). This supports the standard four-class patch label set

$\mathcal{Y} = \{\text{NC}, G3, G4, G5\},$

which is the label space used by multiple later classification and safety studies (Mohammadi-Seif et al., 4 May 2026, Mohammadi-Seif et al., 14 Apr 2026).

A recurrent secondary formulation is the clinically motivated binary superclass mapping

$$S(y) = \begin{cases} 0 & \text{if } y=\text{NC} \ 1 & \text{if } y \in \{\text{G3}, \text{G4}, \text{G5}\}, \end{cases}$$

where NC is treated as Benign / Safe / Non-Critical and G3–G5 as Malignant / Critical (Mohammadi-Seif et al., 4 May 2026, Mohammadi-Seif et al., 14 Apr 2026). This mapping underlies recent work on clinically dangerous errors, where malignant-to-benign confusions are treated as high-risk failures and intra-malignant confusions are treated as lower-severity disagreements (Mohammadi-Seif et al., 4 May 2026, Mohammadi-Seif et al., 14 Apr 2026).

At the WSI level, label usage diverges. One MIL study formulates SICAPv2 as a 5-class ISUP grading problem from WSI-level annotations, with each WSI treated as a bag of instances and no patch labels used during MIL training (Zaffar et al., 2022). Another mixed-supervision MIL study instead treats slide-level Gleason grading as a multi-label MIL classification task, while deriving noisy instance labels from superpixels and limited pixel-level annotations (Bian et al., 2022). A segmentation study notes that raw SICAPv2 provides pixel-level annotations only for Gleason grades and does not provide explicit benign/background masks, so benign masks are computationally derived before training a four-class segmentation model (Badiezadeh et al., 2024).

Formulation	Label space	Representative studies
Patch classification	$\{\text{NC}, G3, G4, G5\}$	(Silva-Rodríguez et al., 2021, Malekmohammadi et al., 2024, Ahmed, 15 Aug 2025)
Safety superclass analysis	Benign vs Malignant	(Mohammadi-Seif et al., 4 May 2026, Mohammadi-Seif et al., 14 Apr 2026)
WSI-level weak supervision	5-class ISUP or slide-level Gleason labels	(Zaffar et al., 2022, Bian et al., 2022)
Segmentation	Benign plus Gleason-pattern masks after preprocessing	(Badiezadeh et al., 2024)

A common misconception is to treat SICAPv2 as having a single fixed label protocol. The literature shows instead that its annotation richness permits several compatible but non-identical task definitions.

3. Preprocessing conventions and experimental protocols

Preprocessing is highly task dependent. In the foundational system, WSIs are downsampled to $10\times$, tissue regions are identified with an Otsu tissue mask, patches with less than 20\% tissue are discarded, and cancerous-biopsy patches with no cancer annotations are removed (Silva-Rodríguez et al., 2021). This yields the patch-level training set used for end-to-end Gleason grading and cribriform detection (Silva-Rodríguez et al., 2021).

Several later studies standardize input size at $10\times$0, but not identically. Risk-Calibrated Learning resizes all SICAPv2 images to $10\times$1 and uses random rotation, flipping, normalization, together with AdamW, cosine annealing, and ImageNet-pretrained ResNet-50 or ViT-B16 backbones (Mohammadi-Seif et al., 14 Apr 2026). The targeted error-correction study also uses an ImageNet-pretrained ResNet-50 in PyTorch with Adam and learning rate 0.001, trained with standard multi-class cross-entropy over the four classes; its post-hoc GBDT modules operate only on the resulting softmax probability vectors (Mohammadi-Seif et al., 4 May 2026). By contrast, HistoViT applies only a lightweight pipeline—pixel scaling by $10\times$2, channel permutation, resizing to $10\times$3, and batching with batch size 32—and explicitly states that it uses no data augmentation (Ahmed, 15 Aug 2025).

Weakly supervised and segmentation workflows rely on different representations. The mixed-supervision MIL Transformer filters blank areas, applies stain normalization, segments tissue into SLIC superpixels, and extracts 1280-dimensional MobileNetV2 features from $10\times$4 crops around superpixel centroids (Bian et al., 2022). EmbAugmenter processes SICAPv2 WSIs into $10\times$5 patches, embeds them with ImageNet-pretrained ResNet50 into $10\times$6, and trains a DA-GAN on 48,552 $10\times$7 pairs per fold to synthesize embedding-space augmentations (Zaffar et al., 2022). For segmentation, one study first performs stain normalization, then constructs a Tissue-Graph from SLIC superpixels, MobileNetV2 features, and a Region Adjacency Graph, using that pipeline to derive benign masks absent from the original SICAPv2 labels (Badiezadeh et al., 2024).

This heterogeneity suggests that SICAPv2 functions less as a standardized preprocessed benchmark than as a histopathology source corpus instantiated through task-specific pipelines. Consequently, cross-paper metric comparison must be conditioned on preprocessing, label construction, and split protocol.

4. Major methodological uses in the literature

The original automatic grading system on SICAPv2 is built around FSConv, a shallow CNN trained from scratch for four-class patch-wise Gleason prediction, combined with a probability-map reconstruction step and an MLP that maps tissue-grade percentages to biopsy-level primary and secondary Gleason grades (Silva-Rodríguez et al., 2021). The same backbone is then fine-tuned for cribriform pattern detection inside GG4 (Silva-Rodríguez et al., 2021). This established SICAPv2 as a benchmark not only for patch classification but also for end-to-end biopsy scoring.

Weakly supervised learning on SICAPv2 has followed two principal lines. One line uses attention-based MIL, where each WSI is a bag and each patch is an instance; EmbAugmenter augments patch embeddings rather than pixels, using a generator–discriminator architecture in embedding space to reduce computational cost while maintaining MIL performance (Zaffar et al., 2022). A second line uses mixed supervision, in which slide-level labels and superpixel-derived instance labels are jointly optimized within a Transformer. That model uses a learnable class token, instance tokens with 2D sinusoidal positional encoding, and a mixed objective

$10\times$8

with random masking of instances to reduce the effect of inaccurate local labels (Bian et al., 2022).

Recent classification studies evaluate more modern architectures directly on SICAPv2 patches. A comparative benchmark applies YOLOv8x, ViT_base_patch16_224_in21k, and MedMamba / Vision Mamba to four-class patch classification (Malekmohammadi et al., 2024). Another study fine-tunes ViT-Base, patch16, 224 with a four-class linear head and standard cross-entropy, arguing that a relatively vanilla transformer already performs strongly on SICAPv2 under minimal preprocessing (Ahmed, 15 Aug 2025). Segmentation work has likewise used SICAPv2 to compare H-vmunet, SAM, and YOLOv8m, with all models trained on processed multi-class segmentation maps rather than raw Gleason-only masks (Badiezadeh et al., 2024).

Safety-aware learning has become a separate methodological theme. Risk-Calibrated Learning introduces a confusion-aware severity matrix $10\times$9 over the four SICAPv2 classes, with $512 \times 512$0 for benign-to-malignant confusions and $512 \times 512$1 for malignant-to-benign confusions, so that fatal false negatives are penalized more heavily than intra-malignant confusions (Mohammadi-Seif et al., 14 Apr 2026). A complementary post-hoc approach applies a dual-classifier GBDT pipeline to the ResNet-50 softmax outputs: one GBDT predicts whether a base prediction is correct, and a second predicts whether an error is human-like or non-human, with only predicted non-human errors being corrected by a forced superclass flip (Mohammadi-Seif et al., 4 May 2026). A related robustness line uses Uncertainty-Aware Multi-Expert Knowledge Distillation, aligning shallow and compact features and weighting decoupled distillation terms by expert uncertainty under source-imbalanced and target-imbalanced SICAPv2 settings (Tong et al., 1 May 2025).

5. Reported empirical performance across tasks

The original SICAPv2 end-to-end system reports patch-level quadratic Cohen’s kappa of 0.77 on the held-out test subset for four-class Gleason grading, AUC of 0.82 for cribriform pattern detection, and quadratic Cohen’s kappa of 0.81 for biopsy-level Gleason scoring (Silva-Rodríguez et al., 2021). Within that work, the shallow FSConv + GMP configuration is emphasized because it reaches performance comparable to or better than fine-tuned state-of-the-art architectures while using far fewer trainable parameters (Silva-Rodríguez et al., 2021).

Mixed-supervision MIL achieves stronger slide-level metrics. The Transformer-based model with spatial encoding and 50\% masking reports macro AUC $512 \times 512$2 on SICAPv2, exceeding both slide-only and prior mixed-supervision baselines in that evaluation protocol (Bian et al., 2022). In another WSI-level MIL setting, EmbAugmenter improves on no augmentation and approaches patch-space augmentation: no augmentation gives 46.5 $512 \times 512$3 accuracy, patch augmentation gives 52.3 $512 \times 512$4, and EmbAugmenter (Ind) also gives 52.3 $512 \times 512$5, while embedding-space augmentation is reported to be more than 300× faster computationally than pixel-space augmentation (Zaffar et al., 2022).

Patch-level architectural comparisons vary sharply with setup. On the four-class classification benchmark using 10,340 patches, MedMamba reports Precision 64.61, Recall 58.23, F1-score 61.25, and Accuracy 69.98, compared with ViT Accuracy 66.02 and YOLO Accuracy 64.70 (Malekmohammadi et al., 2024). In the ViT-only HistoViT study, the reported SICAPv2 results are much higher—Accuracy 96.92\%, Precision 96.91\%, Recall 96.92\%, and AUC 99.71—under that paper’s own patch set and split protocol (Ahmed, 15 Aug 2025). These numbers should not be read as a single leaderboard because the underlying sample counts, splits, and preprocessing are not the same across papers.

SICAPv2 has also been used beyond classification. For segmentation, H-vmunet reports DSC (F1) $512 \times 512$6, Precision $512 \times 512$7, and Recall $512 \times 512$8 on SICAPv2, exceeding SAM and YoloS in that study (Badiezadeh et al., 2024). For retrieval, the unsupervised CAE-based UCBMIR system reports top-5 recall 80\% under its first evaluation protocol and precision 70\% under its second protocol on SICAPv2, while using a 200-dimensional latent representation learned without Gleason labels during feature learning (Tabatabaei et al., 2023).

6. Safety, robustness, and unresolved issues

SICAPv2 has recently become a benchmark for clinically weighted error analysis. In Risk-Calibrated Learning, the primary safety metric is the Critical Error Rate (CER), defined as the proportion of malignant samples misclassified as benign. On SICAPv2, with ResNet-50, Focal Loss yields CER 9.07\%, F1-Macro 0.561, and Accuracy 63.6\%, whereas RCL-Proposed yields CER 6.23\%, F1-Macro 0.599, and Accuracy 69.3\%; with ViT-B16, RCL reduces CER from 10.69\% to 0.81\% but also reduces F1-Macro from 0.666 to 0.436 and Accuracy from 71.1\% to 57.8\% (Mohammadi-Seif et al., 14 Apr 2026). The SICAPv2 results therefore illustrate both a strong safety gain and a possible safety–accuracy trade-off, depending on architecture.

The post-hoc targeted error-correction framework studies a different safety notion: the rate of benign–malignant superclass confusions. On SICAPv2, the base ResNet-50 has class accuracy 68.43\%, super-class accuracy 91.00\%, and 191 non-human errors among 2,122 test patches; after the dual-GBDT pipeline and superclass-flip policy, these become class accuracy 69.18\%, super-class accuracy 92.13\%, and 167 non-human errors, corresponding to a 12.57\% reduction in dangerous non-human errors (Mohammadi-Seif et al., 4 May 2026). The same study reports 0.1 ms per-image overhead and 1.70\% relative latency overhead on SICAPv2, indicating low post-hoc deployment cost (Mohammadi-Seif et al., 4 May 2026).

Imbalance and domain shift are additional recurrent concerns. UMKD explicitly notes that, in SICAPv2, stage III prostate cancer samples constitute 8\% of the cohort, and reports state-of-the-art results in both source-imbalanced and target-imbalanced scenarios: in the source-imbalanced setting, OA 91.02\%, mAcc 90.23\%, F1 90.94\%, MAE 0.1294; in the target-imbalanced setting, OA 91.75\%, mAcc 90.72\%, F1 91.72\%, MAE 0.1199 (Tong et al., 1 May 2025). Other papers emphasize related limitations: limited and partly inaccurate instance labels in mixed-supervision MIL (Bian et al., 2022), lack of explicit benign masks in raw segmentation annotations (Badiezadeh et al., 2024), and clinically significant inter-observer variability in Gleason grading and cribriform recognition (Silva-Rodríguez et al., 2021).

A persistent interpretive issue is that SICAPv2 supports multiple valid benchmark constructions but not a universally standardized one. Reported corpus sizes, split rules, patch definitions, and label abstractions differ substantially across studies. This suggests that direct method ranking without harmonized protocol is methodologically unsafe. At the same time, the breadth of these uses explains the benchmark’s continuing importance: SICAPv2 remains one of the few prostate pathology resources simultaneously used for patch classification, slide-level weak supervision, segmentation, retrieval, safety-aware optimization, and post-hoc reliability correction (Silva-Rodríguez et al., 2021, Mohammadi-Seif et al., 4 May 2026).