METU CCTGS: CRC Tumor Segmentation Data

• METU CCTGS dataset is a public digital histopathology resource featuring 103 whole-slide images with expert pixel-level annotations across five tissue classes.
• It provides both high-resolution WSIs and downsampled mini-WSIs, enabling standardized evaluation of automated colorectal cancer grading under varying computational constraints.
• The dataset facilitates benchmarking semantic segmentation methods with defined metrics like macro F-score and mIoU, thus guiding improvements in tumor grading precision.

Searching arXiv for the specified dataset paper and closely related colorectal histopathology segmentation work. The METU Colorectal Cancer Tumor Grade Segmentation (CCTGS) dataset is a publicly available digital histopathology resource for semantic segmentation of colorectal cancer (CRC) tissue on whole-slide images (WSIs). It was introduced in connection with the ICIP Grand Challenge on Colorectal Cancer Tumor Grading and Segmentation and is designed to support automated, standardized analysis of CRC grading, a task that the source paper characterizes as clinically important yet subjective and prone to observer variability. The dataset comprises 103 WSIs with expert pixel-level annotations for five tissue classes, and it is distributed in both original whole-slide form and a downsampled “mini-WSI” form to accommodate different compute capacities (Bahcekapili et al., 7 Jul 2025).

1. Clinical and methodological scope

CRC is described as the third most diagnosed cancer and the second leading cause of cancer-related death worldwide. Within that setting, histopathological grading is essential for prognosis and treatment planning. The dataset is explicitly motivated by the need to reduce subjectivity in grading and to mitigate constraints associated with global shortages of trained pathologists (Bahcekapili et al., 7 Jul 2025).

Methodologically, METU CCTGS is framed as a benchmark for semantic segmentation rather than slide-level classification or retrieval. Its target output is a dense tissue mask over histopathology imagery, with grading information embedded at the pixel level. This distinguishes it from weakly supervised pathology corpora and from image-level diagnostic datasets. A plausible implication is that the dataset is intended not only for region detection but also for fine-grained morphologic discrimination among tumor grades.

The “From Giga to Mini Challenge” framing reflects the scale transition from native-resolution gigapixel WSIs to downsampled representations. This suggests a dual-use design: evaluation on realistic whole-slide pathology imagery and accessibility for groups with more limited GPU memory or storage budgets.

2. Image corpus, annotation protocol, and label space

The dataset comprises $N = 103$ digitized WSIs, with one slide per patient. The slides were originally acquired at high optical magnifications, specifically $20\times$ or $40\times$, and stored as SVS files. Native resolutions reach several tens of thousands of pixels per dimension, with the source text giving $100\,000 \times 50\,000$ as an example. Each WSI is also provided in a downsampled “mini-WSI” form, typically using $4\times$ downsampling, to accommodate different compute capacities (Bahcekapili et al., 7 Jul 2025).

Five mutually exclusive tissue classes were defined by expert gastrointestinal pathologists.

Class	Description
Grade 1 tumor	well-differentiated
Grade 2 tumor	moderately differentiated
Grade 3 tumor	poorly differentiated
Normal mucosa	non-tumor mucosa
Others	e.g., stroma, artifacts, lymphoid aggregates

Pixel-level annotations were created by manual delineation of each class boundary on the high-resolution SVS files using QuPath, with the source explicitly citing Bankhead et al., Scientific Reports (2017). Two stylized annotation sets are present: one with smoothed boundaries and one capturing highly jagged edges. The stated purpose is to reflect real-world variability in pathologist markings (Bahcekapili et al., 7 Jul 2025).

All slides underwent a visual quality-control pass to exclude regions with gross artifacts, specifically folds and air bubbles, and to ensure label consistency across the two annotation styles. No stain-normalization was enforced. However, the documentation states that users may apply methods such as Macenko or Reinhard normalization as part of their own pipelines. For training convenience, intensity scaling to $[0,1]$, per-channel mean subtraction, and standard deviation normalization were applied to each $512 \times 512$ or $1\,024 \times 1\,024$ patch. Taken together, these design choices indicate that the dataset preserves substantial freedom in preprocessing while still defining a practical baseline input protocol.

3. Partitioning strategy and class distribution

The 103 WSIs were stratified by overall tissue content at the pixel level and then randomly partitioned into training, validation, and test sets. The split is:

• Training: 72 slides (70%)
• Validation: 15 slides (15%)
• Test: 16 slides (15%)

Within each split, the relative pixel proportions of the five classes remain approximately constant, so that no class is under- or over-represented in validation or test (Bahcekapili et al., 7 Jul 2025).

The average class composition across annotated pixels is reported as approximately 40% normal mucosa, approximately 45% combined tumor grades 1–3, and approximately 15% “others.” Because the tumor content is further subdivided into three grade classes, the task remains class-sensitive even if the aggregate tumor fraction is substantial. This suggests that the main statistical challenge is not merely tumor-versus-normal separation, but the separation of morphologically distinct tumor grades under realistic whole-slide heterogeneity.

The split construction matters operationally. Stratification by pixel-level tissue content reduces the likelihood that validation or test performance is dominated by atypical slides with skewed tissue composition. For benchmarking, this supports more stable comparison across segmentation models and postprocessing strategies.

4. Evaluation protocol and metrics

Challenge submissions were made through Codalab and evaluated on a hidden test set using two primary metrics computed per class $c$ and then averaged over the $C = 5$ classes (Bahcekapili et al., 7 Jul 2025).

Let $20\times$0, $20\times$1, and $20\times$2 denote true positives, false positives, and false negatives for class $20\times$3. The reported metrics are:

$20\times$4

and

$20\times$5

The use of macro averaging across the five classes is consequential. It weights all classes equally at evaluation time rather than by pixel frequency. A plausible implication is that improvements on less prevalent or more difficult classes, such as specific tumor grades, can materially affect leaderboard standing even when those classes occupy fewer pixels than normal mucosa.

The evaluation design is also tightly aligned with dense prediction. Unlike slide-level grading accuracy, these metrics explicitly reward correct localization and penalize both omission and over-segmentation at class level. The presence of both $20\times$6 and mIoU further supports comparison between methods that may trade boundary sensitivity against region overlap.

5. Baselines and challenge outcomes

The reference baseline was a patch-based Swin Transformer pretrained on ImageNet and fine-tuned on the downsampled training slides. In the detailed benchmark description, it achieved a macro F-score of 62.9%, with precision approximately 60.9% and recall approximately 69.6%. The challenge abstract reports the baseline as 62.92 F-score. Among 39 participating teams, six methods outperformed this baseline on the hidden test set after Docker-verified runs (Bahcekapili et al., 7 Jul 2025).

Method	Macro F-score	mIoU
VAN+UperNet	70.2%	56.5%
DPT+MaxViT	69.8%	55.8%
HardNet+Lawin	66.7%	52.8%
Segmenter-L	65.2%	51.3%
SegFormer-B1	65.1%	50.7%
PathVTA	64.2%	50.2%

The leading submission, VAN+UperNet, used a UperNet decoder on a Visual Attention Network backbone, together with an auxiliary FCN head and a combined Dice + cross-entropy loss with label smoothing and background exclusion. Training used three-fold cross-validation on $20\times$7 active-cropped patches sampled by class-weighted region scores, heavy geometric and color augmentations, differential learning rates of $20\times$8 for the encoder and $20\times$9 for the decoder, cosine warmup, and Adam. Inference used overlapping $40\times$0 patches with 50% overlap, Gaussian-weighted fusion of three model outputs, nearest-neighbor interpolation to full test resolution, and contour-based hole filling plus small-region correction per class (Bahcekapili et al., 7 Jul 2025).

The second-ranked DPT+MaxViT system used an ensemble of Dense Prediction Transformer and Multi-Axis ViT encoders of varying scales. It trained on $40\times$1 crops followed by downscaling, used an adaptive augmentation policy optimized via LLM feedback, and mixed loss functions across models. Its inference stack combined hard majority voting, soft Top-$40\times$2 biased voting on class probabilities with bias toward tumor classes, Gaussian smoothing of the probability maps, morphological closing, and removal of small connected components below an area threshold $40\times$3 (Bahcekapili et al., 7 Jul 2025).

The third-ranked HardNet+Lawin system paired a HarDNet encoder pretrained on NCT-CRC-HE-100K classification with a Lawin transformer decoder. It used five-fold cross-validation on downsampled $40\times$4 images, a warm-up unfreezing schedule, standard geometric and color augmentations, and AdamW with cosine annealing and EMA. At inference, it ensembled the top three folds with test-time augmentations and majority voting, followed by hole filling (Bahcekapili et al., 7 Jul 2025).

The remaining top submissions were also informative. Segmenter-L fine-tuned an off-the-shelf Segmenter-L with a ViT backbone from ADE20K-pretrained weights on the CCTGS mini-WSIs. SegFormer-B1 fine-tuned an off-the-shelf SegFormer-B1 with an MiT backbone from Cityscapes-pretrained weights. PathVTA used a frozen UNI foundation model backbone, a ViT Adapter for multi-scale features, and a U-Net-style decoder; inference used sliding-window prediction on $40\times$5 patches with stride 112, accumulation and re-normalization of softmax probabilities, and argmax for the final mask (Bahcekapili et al., 7 Jul 2025).

Across these results, several motifs recur: patch-based training, cross-validation, overlapping-window inference, probability fusion, and morphological cleanup. This suggests that, on METU CCTGS, optimization details and inference engineering are nearly as important as backbone choice.

6. Recommended usage patterns, pitfalls, and implementation considerations

The dataset documentation provides explicit preprocessing recommendations. Active cropping is recommended so that patches are sampled based on semantic richness rather than uniform tiling, with the stated goal of alleviating class imbalance, especially for sparse tumor grades. For multi-center or multi-stain training, stain normalization such as Macenko is recommended to reduce domain shift. Whenever compute permits, training on original WSIs or on large crops of at least $40\times$6 is recommended in order to capture morphological detail essential for distinguishing tumor grades (Bahcekapili et al., 7 Jul 2025).

Postprocessing guidance is equally concrete. Nearest-neighbor interpolation is recommended when discrete label maps are upsampled, in order to avoid class mixing. Gaussian kernels are recommended for patch-level prediction fusion before merging, to reduce edge artifacts. Morphological cleanup should include closing, defined as dilation followed by erosion, together with removal of small connected components below a dataset-specific area threshold (Bahcekapili et al., 7 Jul 2025).

The documented pitfalls center on annotation style variability, ensembling, and resource constraints. Because both smoothed and jagged annotations are present, the source notes that learning can be biased by annotation style. The stated recommendation is to select one annotation domain or to apply morphological smoothing or edging filters uniformly to ground-truth masks. The same documentation reports that voting-based ensembling, whether hard or soft, consistently yielded 3–7 point gains in F-score. Full-resolution WSIs are also described as posing memory challenges, motivating patch-based pipelines with overlap, careful batching, and mixed-precision training (Bahcekapili et al., 7 Jul 2025).

These recommendations indicate that the dataset is not merely a passive benchmark. It is accompanied by a fairly explicit operational doctrine: preserve label discreteness during resizing, reduce boundary artifacts during fusion, manage class imbalance through informed sampling, and treat annotation style as a real source of variance rather than incidental noise.

7. Naming ambiguity and distinction from other METU datasets

An important source of confusion is that “METU” also appears in unrelated computer vision literature. The trademark retrieval paper “Learning Regional Attention over Multi-resolution Deep Convolutional Features for Trademark Retrieval” describes a million-scale METU trademark dataset with nearly 1,000,000 gallery images, 417 query images, 35 visually similar groups, image-level labels only, and evaluation by MAP@100 and NAR; it explicitly reports no bounding boxes and no pixel-level masks (Tursun et al., 2021).

That retrieval corpus is methodologically distinct from METU CCTGS as defined in colorectal histopathology. The histopathology dataset consists of 103 WSIs, one per patient, with expert pixel-level segmentation labels for five tissue classes, stratified train/validation/test splits, and evaluation by macro F-score and mIoU (Bahcekapili et al., 7 Jul 2025). The trademark dataset, by contrast, is a gallery–query retrieval benchmark with group-level relevance labels and no dense annotation (Tursun et al., 2021).

This distinction matters because acronym overlap can obscure dataset identity. In the CRC context, “METU CCTGS” denotes a segmentation benchmark for tumor grading in digital pathology, not a logo or trademark retrieval resource. Within that scope, it functions as a benchmark for dense histopathologic parsing under whole-slide scale, class-specific grading, and annotation-style variability.