MammoGPS

Dataset Summary

MammoGPS is a benchmark for evaluating vision-language model spatial understanding on 2D mammography. The benchmark is designed for analysis-oriented evaluation rather than single-number leaderboard reporting: the goal is to separate failures of generic localization, medically relevant finding recognition, and landmark-grounded spatial reasoning.

This repository currently includes benchmark task views for:

• finding localization
• finding localization with type, quadrant, depth, and fine-depth information
• finding localization with curve/highlight support
• anatomy localization
• finding-type QA
• depth, fine-depth, and quadrant QA with and without landmark aids.
• spatial-bias stress tests based on grid placement and horizontal flipping.

The current repository layout is benchmark-only: there is a single benchmark split and no train/validation/test split.

The current benchmark/images contains processed images from the following datasets:

• cmmd
• cbis_ddsm
• inbreast

vindr dataset requires users to sign data agreement and download the dataself on their own. We provide script to process the downloaded vindr dataset to benchmark specific configuration For details about vindr dataset : https://vindr.ai/datasets/mammo . Dataset download link: https://www.physionet.org/content/vindr-mammo/1.0.0/

These datasets were chosen because they provide annotated clinical findings and their licenses allow derivative use and redistribution. We included only images with a single annotated finding, ensuring that localization, depth, and quadrant labels could be attributed unambiguously to one clinically relevant target.

Exporting VinDr-Mammo dataset

The export script is:

• scripts/export_vindr.py

Example with PNG export:

python3 scripts/export_vindr.py \
  --input-csv scripts/vindr_curated_with_pnl_and_depth_gt_relimpaths.csv \
  --vindr-dir Path to downloaded VinDr-Mammo dataset ( eg: data/physionet.org/files/vindr-mammo/) \
  --output-dir data/benchmark \

CC-specific behavior in the exporter:

• pectoral_muscle excludes CC views
• depth wording uses chest wall for CC where appropriate
• quadrant tasks use outer / inner for CC and upper / lower for MLO

Supported Tasks

Localization tasks

• finding
• finding_with_type
• finding_with_location
• finding_with_depth
• finding_with_fine_depth
• finding_with_type_location_depth
• finding_with_type_location_fine_depth
• finding_with_curve
• finding_with_highlight
• curve_bbox
• highlight_bbox
• curve_bbox_blank
• highlight_bbox_blank
• nipple
• pectoral_muscle

QA tasks

• finding_type
• depth
• depth_pec
• depth_pec_nipple
• depth_pec_nipple_simple
• depth_pec_nipple_simple_blank
• fine_depth
• quadrant
• quadrant_pnl
• quadrant_pnl_simple
• quadrant_pnl_simple_blank

Spatial-bias tasks

• spatial_bias_grid_highlight
• spatial_bias_grid_highlight_blank
• spatial_bias_grid_blank
• spatial_bias_flip_left_finding
• spatial_bias_flip_right_finding
• spatial_bias_flip_left_highlight
• spatial_bias_flip_right_highlight

Benchmark Motivation

Spatial understanding in mammography is clinically meaningful and technically distinct from generic medical VQA. A model may know high-level radiology concepts and still fail on image-grounded localization, quadrant reasoning, or localization relative to mammography-specific landmarks such as the nipple and pectoral muscle. MammoGPS is intended to study these failures in a structured way.

The benchmark includes task families with progressively increasing support, from unconstrained grounding to grounding with type cues, contours, and highlighted regions, plus dedicated stress tests for spatial bias.

Blank-canvas variants are also included for selected contour/highlight tasks so the model can be probed with the finding signal isolated from mammographic background structure.

Dataset Structure

MammoGPS/
├── README.md
├── data/
│   └── benchmark/
│       ├── core.jsonl
│       ├── images/
│       ├── task_*.jsonl
│       └── task_registry.json
└── scripts/
     ├── export_vindr.py
     ├── vindr_curated_with_pnl_and_depth_gt_relimpaths.csv
     ├── eval_qwen3_all_tasks.py
     └── eval_utils.py

The dataset contains a single split:

• benchmark

This is intentional. The repository is designed for benchmarking and analysis rather than model training.

Configurations

Each Hugging Face config in the YAML header maps directly to one task JSONL file in data/benchmark/.

Examples:

• finding -> data/benchmark/task_finding.jsonl
• finding_with_highlight -> data/benchmark/task_finding_with_highlight.jsonl
• depth -> data/benchmark/task_depth.jsonl
• spatial_bias_grid_highlight -> data/benchmark/task_spatial_bias_grid_highlight.jsonl

Data Instances

Each task row contains:

• task_name: task identifier
• sample_id: unique row identifier for the task view
• case_id: benchmark case identifier
• image_id: image identifier
• image_path: either a single image path or, for grid-bias tasks, a 4-item list ordered as [top_left, top_right, bottom_left, bottom_right]
• prompt: the benchmark instruction shown to the model
• response_format: one of xyxy_or_empty, xy_or_empty, or label
• target_text: canonical text answer
• target_bbox_xyxy: canonical ground-truth box for bbox tasks
• target_point_xy: canonical ground-truth point for point tasks
• answer_choices: optional label set for QA tasks
• metadata: subgroup metadata such as dataset_name, view, laterality, finding_type, quadrant, and closest_landmark

The shared data/benchmark/core.jsonl file stores image-level metadata that correspond to the task rows.

Data Fields

Common conventions

• Bounding boxes use absolute image pixel coordinates in xyxy format.
• Negative examples use target_bbox_xyxy = null or target_point_xy = null and target_text = "".
• Prompts are stored directly in the dataset to make evaluation reproducible.
• This Hugging Face-compatible version uses standard data files and YAML config metadata in the dataset card.
• Grid-bias tasks store image components rather than pre-rendered composite PNGs. The evaluation code is expected to compose the 2x2 grid on the fly.
• Flip-bias tasks store the base image path and a metadata.flip_horizontal flag. The evaluation code is expected to flip the image on the fly when this flag is true.
• Exported image paths are dataset-aware, e.g. images/curve/cmmd/... or images/curve/vindr/....
• The exporter appends to existing core.jsonl and task_*.jsonl files rather than overwriting them.

Core metadata fields

The shared core.jsonl table includes:

• case_id
• image_id
• image_path
• dataset_name
• view
• laterality
• width
• height
• has_suspicious_finding
• primary_finding_type
• primary_bbox_xyxy
• quadrant
• closest_landmark

Example Usage

from datasets import load_dataset

ds = load_dataset("mammovlmbench/mammogps", name='finding', split="benchmark")
print(ds[0]["prompt"])
print(ds[0]["target_text"])

If you want the image_path column decoded into an image object:

from datasets import Image, load_dataset

ds = load_dataset("mammovlmbench/mammogps", name='finding', split="benchmark")
ds = ds.cast_column("image_path", Image())
print(ds[0]["image_path"].size)

Grid-bias tasks use list-valued image_path, so they should be rendered by the load_task_image method in benchmark code in scripts/eval_utils.py rather than cast directly with datasets.Image().

Evaluation

Recommended evaluation behavior:

• Bbox tasks: parse model output as x1,y1,x2,y2 or empty string, score negatives by empty prediction and positives by IoU against target_bbox_xyxy.
• Point tasks: parse model output as x,y or empty string and score against target_point_xy.
• Label tasks: score exact match after normalization.
• Report performance overall and by subgroup.

Scripts for evaluation is available within scripts/eval_utils.py folder

Qwen3VL Evaluation scripts:

We provide the evaluation script for Qwen3VL VLM. This could be used as a baseline script and the evaluation part can be extrapolated to other vlms. See scripts/eval_qwen3_all_tasks.py

---

Intended Uses

This dataset is intended for:

• benchmarking vision-language model spatial understanding on 2D mammography, including localization, quadrant reasoning, depth reasoning, laterality, and reasoning relative to mammography-specific landmarks (nipple and pectoral muscle)
• analyzing failure modes in medically grounded localization, where the model must ground its answer in the visual location of a single annotated finding
• studying spatial biases such as laterality sensitivity and image-position sensitivity through controlled stress tests (e.g., horizontally flipped variants, blank-canvas variants, grid-based highlight probes)
• comparing evaluation protocols that disentangle types of model failure across task families, including unconstrained grounding, type-cued grounding, contour-based tasks, highlighted-region tasks, and blank-canvas variants

Out-of-Scope Uses

This dataset is not intended for:

• model training, fine-tuning, or development in its current benchmark-only form (the benchmark contains a single benchmark split with no train/validation/test partition)
• direct clinical decision-making, treatment planning, patient triage, or risk prediction
• standalone cancer screening support, diagnosis, or commercial clinical deployment
• radiology report generation
• measuring general radiology competence outside the defined benchmark tasks
• making claims about real-world clinical sensitivity, specificity, fairness, or robustness without additional external validation
• evaluating multilingual or non-English clinical reasoning (prompts are English-only)

Limitations

• Single-finding restriction. Only images with a single annotated finding are retained, so that localization, depth, quadrant, and landmark-relative labels can be assigned unambiguously. This under-represents normal exams, multi-lesion exams, diffuse findings, and complex real-world presentations.
• Model-derived landmark labels. Nipple and pectoral-muscle annotations come from supervised U-Net detectors trained on an in-house dataset. These labels carry detector uncertainty and may degrade for unusual anatomy, artifacts, poor positioning, weak landmark visibility, post-surgical changes, implants, or acquisition conditions under-represented in the detector training data.
• Evaluation-time image composition. Some task views rely on evaluation-time image composition or flipping rather than fully materialized PNGs. Users should reproduce these transformations exactly as specified in the evaluation code to ensure comparable results.
• No clinical validation. Performance on MammoGPS should not be interpreted as evidence that a model will generalize across all mammography settings, healthcare systems, scanner vendors, acquisition protocols, demographic groups, or clinical workflows.

Bias, Risks, and Ethical Considerations

The benchmark is motivated in part by the possibility of systematic spatial bias in model performance. For example, a model that underperforms on one side of the image could systematically miss findings for a substantial subgroup of cases. The benchmark should therefore be used not only to measure average accuracy, but also to identify subgroup failures and directional biases.

Known sources of bias in MammoGPS include:

• Source-dataset bias. MammoGPS inherits demographic, geographic, equipment, and institutional skews from its four source datasets, which originate from Vietnam (VinDr-Mammo), the United States (CBIS-DDSM), China (CMMD), and Portugal (INbreast). Acquisition technology spans digitized film and several FFDM vendors, so some patient populations, imaging practices, scanners, and clinical settings are over- or under-represented.
• Selection bias from single-finding filtering. Only images with a single annotated finding are included. Normal exams, ambiguous cases, multi-lesion exams, and complex real-world presentations are under-represented.
• Automated-annotation bias. Nipple and pectoral-muscle labels are produced by U-Net detectors trained on an in-house dataset. These labels can inherit detector-specific errors and performance limitations on out-of-distribution anatomy, artifacts, or acquisition conditions.
• Language and task-design bias. Prompts are English-only and tasks are constrained to controlled spatial-reasoning formats, which limits external validity to broader clinical reasoning.

Risks and potential harms. The principal risks are misuse and over-interpretation. MammoGPS should not be treated as evidence of clinical readiness, cancer-detection ability, diagnostic safety, or fairness across patient groups. Specific potential harms from misuse include premature clinical claims, inappropriate model rankings, training or fine-tuning on the benchmark, and unsafe use in screening or diagnostic workflows — any of which could contribute to missed findings, incorrect localization, delayed diagnosis, or unequal performance across patient groups in real clinical settings.

Sensitive information. MammoGPS does not introduce additional patient identifiers and inherits the de-identification practices of the source datasets. Residual privacy risk is expected to be low but non-zero. Sex/gender is indirectly represented (mammograms predominantly come from patients with clinically female anatomy); age, geography, healthcare system, and scanner-device characteristics are indirectly inherited from the source datasets. Socio-economic status, political or religious beliefs, and experience or seniority are not represented.

Licensing

CC-BY-4.0

Citation

TO BE ADDED LATER

Contact

TO BE ADDED LATER