---
dataset_info:
- config_name: findings_section
  features:
  - name: dicom_id
    dtype: string
  - name: study_id
    dtype: string
  - name: subject_id
    dtype: string
  - name: main_image
    dtype: image
  - name: findings_section
    dtype: string
  - name: impression_section
    dtype: string
  - name: indication_section
    dtype: string
  - name: comparison_section
    dtype: string
  - name: technique_section
    dtype: string
  - name: history_section
    dtype: string
  - name: examination_section
    dtype: string
  - name: prior_image
    dtype: image
  - name: default_prompt
    dtype: string
  - name: PerformedProcedureStepDescription
    dtype: string
  - name: ViewPosition
    dtype: string
  - name: Rows
    dtype: int32
  - name: Columns
    dtype: int32
  - name: StudyDate
    dtype: string
  - name: StudyTime
    dtype: string
  - name: ProcedureCodeSequence_CodeMeaning
    dtype: string
  - name: ViewCodeSequence_CodeMeaning
    dtype: string
  - name: PatientOrientationCodeSequence_CodeMeaning
    dtype: string
  - name: view
    dtype: string
  splits:
  - name: test
    num_bytes: 7996686675.113
    num_examples: 2461
  download_size: 7519081766
  dataset_size: 7996686675.113
- config_name: impression_section
  features:
  - name: dicom_id
    dtype: string
  - name: study_id
    dtype: string
  - name: subject_id
    dtype: string
  - name: main_image
    dtype: image
  - name: findings_section
    dtype: string
  - name: impression_section
    dtype: string
  - name: indication_section
    dtype: string
  - name: comparison_section
    dtype: string
  - name: technique_section
    dtype: string
  - name: history_section
    dtype: string
  - name: examination_section
    dtype: string
  - name: prior_image
    dtype: image
  - name: default_prompt
    dtype: string
  - name: PerformedProcedureStepDescription
    dtype: string
  - name: ViewPosition
    dtype: string
  - name: Rows
    dtype: int32
  - name: Columns
    dtype: int32
  - name: StudyDate
    dtype: string
  - name: StudyTime
    dtype: string
  - name: ProcedureCodeSequence_CodeMeaning
    dtype: string
  - name: ViewCodeSequence_CodeMeaning
    dtype: string
  - name: PatientOrientationCodeSequence_CodeMeaning
    dtype: string
  - name: view
    dtype: string
  splits:
  - name: test
    num_bytes: 7553630749.877
    num_examples: 2343
  download_size: 7149201197
  dataset_size: 7553630749.877
configs:
- config_name: findings_section
  data_files:
  - split: test
    path: findings_section/test-*
- config_name: impression_section
  data_files:
  - split: test
    path: impression_section/test-*
task_categories:
- image-text-to-text
language:
- en
tags:
- chest-xray
- report-generation
- mimic-cxr
pretty_name: MIMIC-CXR Radiology Report Generation
size_categories:
- 1K<n<10K
---

# MIMIC-CXR-RRG: Radiology Report Generation Subsets

This dataset provides two carefully filtered and structured subsets from the [MIMIC-CXR](https://physionet.org/content/mimic-cxr/2.0.0/) dataset, specifically designed for **Radiology Report Generation (RRG)** tasks. It includes image-report pairs focused on the **Findings** and **Impression** sections, targeting **frontal-view chest X-rays** only.

---

## 📚 Dataset Overview

| Subset              | Section Target     | Split | #Samples | View Type     |
|---------------------|--------------------|-------|----------|----------------|
| `findings_section`  | Findings            | test  | 2361     | Frontal only   |
| `impression_section`| Impression          | test  | 2343     | Frontal only   |

- The splits **follow the evaluation protocol** used in models such as [Libra](https://x-izhang.github.io/Libra_v1.0/) and [MAIRA-2](https://arxiv.org/abs/2406.04449).
- Images and labels are provided in a test-only setting, useful for **benchmarking and zero-shot evaluation**.

---

## 🧾 Data Format

Each instance in both subsets contains:

- 📷 `main_image` – The frontal-view chest X-ray
- 📷 `prior_image` – (Optional) Prior image if available
- 📝 Text sections:
  - `findings_section`
  - `impression_section`
  - `indication_section`
  - `comparison_section`
  - `technique_section`
  - `history_section`
  - `examination_section`
- 💬 `default_prompt` – Prompt for generation tasks
- 🧾 Metadata:
  - `dicom_id`, `study_id`, `subject_id`
  - Acquisition info: `Rows`, `Columns`, `StudyDate`, `ViewPosition`, etc.

---

## 🚀 How to Use

```python
from datasets import load_dataset

# Load a specific subset (e.g., findings_section)
ds = load_dataset("X-iZhang/MIMIC-CXR-RRG", name="findings_section", split="test")

# Display an image
from PIL import Image
ds[0]["main_image"].show()

# View sample
print(ds[0]["findings_section"])
```

## ✏️ Citation
```
@misc{zhang2025libraleveragingtemporalimages,
      title={Libra: Leveraging Temporal Images for Biomedical Radiology Analysis}, 
      author={Xi Zhang and Zaiqiao Meng and Jake Lever and Edmond S. L. Ho},
      year={2025},
      eprint={2411.19378},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.19378}, 
}
```