Grounding Medical AI in Expert‑Labeled Data: A Case Study on PadChest-GR- the First Multimodal, Bilingual, Sentence‑Level Dataset for Radiology Reporting

This breakthrough in multimodal radiology is a game changer

Introduce yourself

Recent developments in AI-based medical diagnosis have demonstrated that success is dependent not only on sophisticated models, but also on the depth and quality of the data. The case study below highlights an innovative collaboration between Centaur.aiMicrosoft Research and University of Alicante culminating in PadChest‑GR—the first multimodal, bilingual, sentence‑level dataset for grounded radiology reporting. By aligning structured clinical text with annotated chest‑X‑ray imagery, PadChest‑GR empowers models to justify each diagnostic claim with a visually interpretable reference—an innovation that marks a critical leap in AI transparency and trustworthiness.

Moving beyond image classification is a challenge.

HistThe following are some examples of how to useically, medical imaging datasets have supported only image‑level classification. For example, an X‑ray might be labeled as “showing cardiomegaly” or “no abnormalities detected.” Such classifications, while functional and useful, lack in explanation and accuracy. AI models that are trained this way can be prone to misunderstandings. Hallucinations—generating unsupported findings or failing to localize pathology accurately  .

Enter Reporting on ground radiology. This approach demands a richer, dual‑dimensional annotation:

• Space groundingLocalization of findings is indicated by bounding boxes in the image.
• Language groundingEach description text is more specific than a generic classification.
• Contextual clarificationEvery report entry has been contextualized in both the language and spatial sense, which reduces ambiguity while increasing interpretability.

This paradigm shift requires a fundamentally different kind of dataset—one that embraces complexity, precision, and linguistic nuance.

Human‑in‑the‑Loop at Clinical Scale

Creating PadChest‑GR required uncompromising annotation quality. Centaur.ai’s HIPAA‑compliant labeling platform The University of Alicante has trained radiologists to perform:

• Draw bounding boxes around visible pathologies in thousands of chest X‑rays.
• Link each region to specific sentence‑level findings, in both Spanish and English.
• Conduct rigorous, consensus‑driven quality control, including adjudication of edge cases and alignment across languages.

Centaur.ai’s platform is purpose‑built for medical‑grade annotation workflows. Some of its most notable features are:

• Multiple annotator consensus & disagreement resolution
• Performance‑weighted labeling Expert annotations weighted by historical agreement
• Support for The DICOM format and other medical imaging formats
• Workflows multimodal This software handles images, texts, and clinical metadata
• The Full Story Audit Trails, version control, and live quality monitoring—for traceable, trustworthy labels  .

This allowed the team of researchers to focus their efforts on difficult medical details without losing speed or accuracy.

The Dataset: PadChest‑GR

PadChest‑GR builds on the original PadChest dataset by adding these robust dimensions of spatial grounding and bilingual, sentence‑level text alignment  .

Key Features

• Multimodal: Integrates image data (chest X‑rays) with textual observations, precisely aligned.
• BilingualThe annotations are captured in both languages Spanish and EnglishThe broader utility of the product and its inclusivity.
• Sentence‑level granularityRather than a label, each finding has a specific phrase attached to it.
• Visual ExplainabilityThe model is able to show where exactly a diagnostic has been done, which promotes transparency.

By combining these attributes, PadChest‑GR stands as a landmark dataset—reshaping what radiology‑trained AI models can achieve.

Results and Implications

Enhanced Interpretability & Reliability

The models can be positioned to pinpoint the precise region that prompted a particular finding. This greatly improves transparency. Clinicians can see both the claim and its spatial basis—boosting trust.

Reduced AI hallucinations

By tying linguistic claims to visual evidence, PadChest‑GR greatly diminishes the risk of fabricated or speculative model outputs.

Multilingual utility

Multilingual annotations extend the dataset’s applicability across Spanish‑speaking populations, enhancing accessibility and global research potential.

Scalable, High‑Quality Annotation

A combination of expert radiologists with a strict consensus and a secured platform enabled the team to create complex multimodal annotations on scalable basis, without compromising on quality.

Wider Reflections on Why Data Matters for Medical AI

This case study is an illuminating testimony to a larger truth. The future of AI is dependent on data and not models.  . AI is only as good as its foundation, especially in the healthcare industry, where trust and high stakes are at play.

The success of PadChest‑GR hinges on the synergy of:

• Domain experts Radiologists who can make a nuanced judgement.
• Advanced annotation infrastructure (Centaur.ai‘s platform) enabling traceable, consensus-driven workflows.
• Collaborative partnerships Microsoft Research, University of Alicante, and other institutions involved in ensuring technical, scientific and linguistic rigor.

Centaur.ai: Case Study Contextualized

This study is a good example of how to apply the principles in radiology. Centaur.ai‘s wider mission: to scale expert‑level annotation for medical AI across modalities.

• They are able to do this by using their DiagnosUs app, Centaur Labs (the same organization) has built a gamified annotation platform, harnessing collective intelligence and performance‑weighted scoring to label medical data at scale, with speed and accuracy  .
• Their platform is HIPAA‑ and SOC 2‑compliant, supporting annotators across image, text, audio, and video data—and serving clients such as Mayo Clinic spin‑outs, pharmaceutical firms, and AI developers  .
• Innovations like performance‑weighted labeling help ensure that only high‑performing experts influence the final annotations—raising quality and reliability  .

PadChest‑GR sits squarely within this ecosystem—leveraging Centaur.ai’s sophisticated tools and rigorous workflows to deliver a groundbreaking radiology dataset.

The conclusion of the article is:

The PadChest‑GR case study exemplifies how expert‑grounded, multimodal annotation can fundamentally transform medical AI—enabling transparent, reliable, and linguistically rich diagnostic modeling.

By harnessing domain expertise, multilingual alignment, and spatial grounding, Centaur.ai, Microsoft Research, and the University of Alicante have set a new benchmark for what medical image datasets can—and should—be. The success of their project highlights the fact that AI’s promise in the healthcare sector is limited by the quality of the datasets it has been trained with.

This case stands as a compelling model for future medical AI collaborations—highlighting the path forward to trustworthy, interpretable, and scalable AI in the clinic.  Visit for more information. Centaur.ai.

Thanks to the Centaur.ai team for the thought leadership/ Resources for this article. Centaur.ai The team is sponsoring and supporting this article/content.