GADVR: Gastric Adenocarcinoma Visual Reasoning Benchmark

• The paper introduces a dataset that couples patch-level imagery, pixel-level nuclei masks, and image-grounded Q&A pairs to enable interpretable diagnostic reasoning.
• GADVR is constructed via a multi-stage pipeline using CLIP-ViT filtering, HoverNet segmentation, and GPT-4o-generated queries, ensuring detailed and cell-specific annotations.
• The benchmark supports reasoning segmentation, referring segmentation, and image-grounded conversation, providing a comprehensive tool for evaluating multimodal visual reasoning in pathology.

Searching arXiv for the cited papers to ground the article in current preprints. GADVR is a gastric pathology benchmark introduced with PathMR as, to the authors’ knowledge, the first pixel-level multimodal visual reasoning resource tailored to pathology (Zhang et al., 28 Aug 2025). It is designed for H&E-stained gastric adenocarcinoma histopathology and couples cell-level nuclei masks with image-grounded question–answer pairs so that models can jointly localize diagnostically relevant cell populations, reason over morphology and spatial distribution, and produce expert-style diagnostic explanations. In the PathMR framework, GADVR functions both as a training corpus and as an evaluation benchmark for interpretable pathology systems that must align textual rationales with pixel-grounded visual evidence (Zhang et al., 28 Aug 2025).

1. Position within pathology data resources

GADVR was created to address a specific gap between two established dataset regimes in computational pathology (Zhang et al., 28 Aug 2025). On one side are large multimodal pathology collections that provide only slide- or patch-level labels or captions. On the other side are nuclei-level datasets that are rich in pixel-wise annotations but too small to train and benchmark modern vision–language reasoning systems. GADVR explicitly targets this missing middle: a large-scale dataset with patch-level imagery, pixel-level nuclei supervision, and paired language grounded in those pixel annotations.

The dataset is specialized to the clinical setting of gastric cancer, focusing on H&E-stained histopathology from gastric adenocarcinoma. Its design objective is interpretable, cell-level reasoning that mirrors how pathologists diagnose disease—by jointly localizing cell populations, considering their morphology and spatial distribution, and articulating those findings as diagnostic explanations (Zhang et al., 28 Aug 2025). A plausible implication is that GADVR is intended not merely for classification, but for multimodal evidence synthesis in which segmentation and narrative explanation are evaluated together.

2. Source corpus and construction pipeline

GADVR is derived from PatchGastricADC22, a large patch repository for gastric adenocarcinoma subtype diagnosis (Zhang et al., 28 Aug 2025). PatchGastricADC22 contains 262,777 image patches at $40\times$ magnification, each of size $600 \times 600$ pixels, with nine gastric adenocarcinoma subtypes assigned from the source WSI labels. To improve the local relevance of patches to their subtypes, the authors first filter out weakly correlated patches using a pretrained CLIP-ViT with the original diagnostic reports; the retained high-correlation subset is then used to construct GADVR.

The construction process therefore includes three major stages: patch filtering, nuclei annotation generation, and language-pair generation. Figure 1 of the paper summarizes this pipeline as CLIP-ViT filtering against WSI diagnostic reports, HoverNet-based nuclei segmentation and classification, and GPT-4o question/answer generation conditioned on both the image and the masks (Zhang et al., 28 Aug 2025).

Several acquisition details are explicitly left unspecified in the paper. Scanner manufacturers and microns-per-pixel are not reported. All imagery is standard H&E histopathology at $40\times$ in patch form rather than full WSIs, and no explicit color normalization steps are described (Zhang et al., 28 Aug 2025). This suggests that reproducibility across institutions or staining pipelines may require additional downstream harmonization not formalized in the benchmark specification.

3. Annotation layers and language grounding

GADVR provides pixel-level nuclei annotations with both instance masks and cell-type labels (Zhang et al., 28 Aug 2025). Nuclei segmentation and classification are generated using HoverNet pretrained on PanNuke, selected after experts compared models trained on ConSeP, Lizard, and PanNuke and found that PanNuke-pretrained HoverNet achieved superior segmentation quality for the gastric domain. To raise classification fidelity, two pathologists manually corrected misclassified nuclei for approximately 200 patches per category and fed those corrections into a human-in-the-loop refinement cycle.

Quality control is multi-expert. Three independent pathologists performed external quality control on a random sample of 200 images, jointly reviewing both nuclei segmentation and the paired QA annotations. In the paper’s quality summary, over 92% of segmentation masks were judged Correct, and more than 95% of QA pairs were rated Okay or Good across assessors (Zhang et al., 28 Aug 2025). Inter-annotator agreement statistics are not reported, but the dataset incorporates iterative corrections and external review as practical quality assurance. The semantic cell classes used for evaluation are neoplastic (tumor), inflammatory, connective, and epithelial cells.

The textual modality is constructed to be cell-grounded rather than merely patch-descriptive. The authors curate a pool of roughly 200 diagnostic questions using GPT-4o, divided into two complementary families: general morphology questions, covering nuclear morphology, spatial distribution, and tissue architecture, and subtype-specific questions, eliciting diagnostic reasoning about gastric adenocarcinoma subtypes inferred from cellular features (Zhang et al., 28 Aug 2025). For each image, three questions are sampled from this pool and answered by GPT-4o.

Crucially, GPT-4o is prompted with both the selected question or questions and the corresponding nuclei segmentation masks, so that answers explicitly incorporate cell categories and spatial distribution patterns visible in the image (Zhang et al., 28 Aug 2025). The dataset does not claim explicit span-to-instance alignment labels such as token-to-nucleus links, but its question–answer pairs are created conditionally on the masks to ensure strong, cell-level grounding of language to visual content.

4. Scale, splits, and enabled tasks

After CLIP-ViT filtering and generation of nuclei labels and QA pairs, GADVR contains approximately 190,000 pathology image patches and over 547,000 image–text pairs (Zhang et al., 28 Aug 2025). It inherits the nine gastric adenocarcinoma subtype labels from PatchGastricADC22, including examples such as well-differentiated tubular adenocarcinoma, moderately differentiated tubular adenocarcinoma, and papillary adenocarcinoma, although per-class counts are not reported.

The dataset is split by whole slide image to avoid leakage, using an $8{:}1{:}1$ train/validation/test partition. All reported results are WSI-disjoint but intra-dataset; no cross-institutional or external validation split is described (Zhang et al., 28 Aug 2025).

Split	Patches	Image–text pairs
Training	149,928	447,170
Validation	14,814	44,080
Testing	18,754	56,067

At the data-organization level, each patch is $600 \times 600$ at $40\times$, paired with nuclei segmentation and classification labels and a set of three QA pairs sampled from a pool of roughly 200 questions (Zhang et al., 28 Aug 2025). Text is paired to images at the patch level, and answers are conditioned on the masks to ensure grounding in cell categories and spatial distributions.

GADVR is expressly designed to support three tasks. The first is reasoning segmentation, in which image and question are given as input and the model must produce both a textual explanation and pixel-level segmentation masks of the referenced cell types. The second is referring segmentation, in which a textual reference to a cell category or pattern must be grounded by segmenting the corresponding regions. The third is conversation, namely image-grounded Q&A about diagnostic, morphology, and subtype-related content (Zhang et al., 28 Aug 2025). Figure 2 provides representative examples of all three task types, including prompts such as “segment inflammatory cells” and rationales discussing tumor, stromal, and inflammatory composition.

For practical use, the benchmark does not require preprocessing beyond loading the $600 \times 600$ H&E patches (Zhang et al., 28 Aug 2025). The authors’ construction pipeline applied CLIP filtering and HoverNet-generated nuclei labels, but users of GADVR do not need to repeat those steps. Models are trained and evaluated at the native $600 \times 600$ resolution in the reported experiments. Common optimizer settings such as AdamW are mentioned as typical, while data augmentations are not detailed.

5. Evaluation protocol and benchmark behavior

For segmentation, GADVR adopts grounded IoU ($gIoU$) and class IoU ($cIoU$) following prior pixel-level reasoning work (Zhang et al., 28 Aug 2025). Both reduce to standard intersection-over-union at mask level:

$600 \times 600$0

Class IoU averages IoU over classes:

$600 \times 600$1

where $600 \times 600$2 and $600 \times 600$3 are the ground-truth and predicted masks for class $600 \times 600$4. Grounded IoU aggregates IoU over grounded segments predicted in response to the query and is reported alongside $600 \times 600$5 to characterize overall and per-class performance (Zhang et al., 28 Aug 2025). Pixel accuracy is not used.

For the conversation task, the reported metrics are BLEU-4 and token-level $600 \times 600$6 (Zhang et al., 28 Aug 2025). BLEU-4 follows the standard brevity-penalized geometric mean of 1–4 gram precisions, and $600 \times 600$7 is computed from precision $600 \times 600$8 and recall $600 \times 600$9 as

$40\times$0

ROUGE, METEOR, CIDEr, BERTScore, CLIPScore, and retrieval-style metrics are not used on GADVR in this paper.

Although GADVR is a dataset rather than a model, the paper also specifies a benchmark training recipe that clarifies the losses tied to its tasks (Zhang et al., 28 Aug 2025). The weighted mask loss combines BCE and Dice:

$40\times$1

with $40\times$2 on positives, $40\times$3, and $40\times$4. The Dice coefficient is

$40\times$5

and a standard form of weighted BCE is

$40\times$6

A spatial consistency constraint is introduced to reduce morphological fragmentation by encouraging local smoothness across 4-neighborhoods:

$40\times$7

The text-generation loss is standard autoregressive cross-entropy:

$40\times$8

The total loss is

$40\times$9

with all $8{:}1{:}1$0 set to 1 by default (Zhang et al., 28 Aug 2025).

Benchmarking on GADVR includes pixel-grounded visual reasoning baselines fine-tuned from natural-image settings—OV-Seg, LISA (7B, 13B), PixelLM, GSVA, and MMR (7B, 13B)—as well as biomedical assistants for conversation and referring comparisons, namely LLaVA-Med and BiomedParse (Zhang et al., 28 Aug 2025). On reasoning segmentation, PathMR-7B achieves $8{:}1{:}1$1 $8{:}1{:}1$2 and $8{:}1{:}1$3 $8{:}1{:}1$4 on the test set, exceeding MMR-7B at $8{:}1{:}1$5. PathMR-13B reaches $8{:}1{:}1$6 $8{:}1{:}1$7 and $8{:}1{:}1$8 $8{:}1{:}1$9. On referring segmentation, PathMR-13B attains $600 \times 600$0 $600 \times 600$1 and $600 \times 600$2 $600 \times 600$3 overall, surpassing MMR-13B at $600 \times 600$4. On image-grounded conversation, PathMR reports BLEU-4 $600 \times 600$5 and $600 \times 600$6 on test, outperforming LLaVA-Med at $600 \times 600$7 and also outperforming LISA-13B (Zhang et al., 28 Aug 2025). The paper further reports per-category segmentation results for neoplastic, inflammatory, connective, and epithelial cells, noting strong or best performance across categories, including the rarer epithelial class where many baselines struggle.

6. Access, limitations, and nomenclature

GADVR is publicly hosted on Hugging Face at https://huggingface.co/datasets/zhangye-zoe/GADVR (Zhang et al., 28 Aug 2025). The paper does not specify the license or access prerequisites, such as data use agreements, and practitioners are directed to the dataset card for licensing terms, citation format, and any usage restrictions. In publications, the authors request citation of the PathMR paper introducing GADVR.

From an ethics and bias perspective, GADVR is derived from an existing de-identified research dataset and augmented with automated segmentation, automated classification, and GPT-4o-generated QA pairs (Zhang et al., 28 Aug 2025). The paper documents external quality control by three independent pathologists, but does not state IRB approvals, consent procedures, scanner or site provenance, microns-per-pixel, or demographic distributions for the source WSIs. Potential biases therefore include institutional and staining variability from the source collection, subtype imbalance across the nine gastric adenocarcinoma subtypes, and GPT-4o biases in language generation. The reported mitigation measures are model selection for nuclei annotation, targeted human corrections of approximately 200 patches per category, and external expert review; broader ethical and demographic bias analyses are not reported.

A recurrent naming issue warrants clarification. GADVR should not be conflated with GAVD, the “Gait Abnormality in Video Dataset” introduced for clinical gait analysis; in that literature, “GADVR” is described as a misprint or informal shorthand for GAVD rather than the official dataset name (Ranjan et al., 2024). In pathology, by contrast, GADVR denotes the Gastric Adenocarcinoma Diagnosis Visual Reasoning dataset introduced with PathMR (Zhang et al., 28 Aug 2025).

Taken together, GADVR defines a large-scale, cell-level, pixel-grounded benchmark for interpretability-focused multimodal reasoning in gastric pathology (Zhang et al., 28 Aug 2025). Its central contribution is not merely added scale, but the explicit coupling of nuclei masks and QA pairs so that models can be evaluated on whether they both localize and verbalize diagnostically salient evidence at the granularity used in pathology practice.