README.md

---
license: cc-by-2.5
task_categories:
- image-classification
- image-segmentation
- image-feature-extraction
language:
- en
tags:
- biology
- X-Ray
size_categories:
- 1K<n<10K
---

# Dataset Card for FracAtlas

<!-- Provide a quick summary of the dataset. -->

The "FracAtlas" dataset is a collection of musculoskeletal radiographs for bone fracture classification, localization, and segmentation. 
It includes 4,083 X-Ray images (717 fracture images) with corresponding annotations in multiple formats. The annotations include segmentations, width, and etc in COCO, VGG, YOLO, and Pascal VOC format. 
The dataset is intended for use in deep learning tasks in medical imaging, specifically targeting the understanding of bone fractures. 
It is freely available under a CC-BY 4.0 license.

This script provides a Hugging Face `datasets` loader for the FracAtlas dataset. The generated dataset includes high-quality X-Ray images and incorporate detailed annotations from COCO JSON format for segmentation
and bounding box information, as well as additional localization data from PASCAL VOC XML files. The loader handles downloading and preparing the dataset, making it readily available for machine learning models and analysis
tasks in medical imaging, especially focusing on the detection and understanding of bone fractures.

- **Curated by:** Abedeen, Iftekharul; Rahman, Md. Ashiqur; Zohra Prottyasha, Fatema; Ahmed, Tasnim; Mohmud Chowdhury, Tareque; Shatabda, Swakkhar
- **License:** cc-by-2.5

### Dataset Sources

<!-- Provide the basic links for the dataset. -->
The source data for the "FracAtlas" dataset is hosted on Figshare, an online digital repository where researchers can preserve and share their research outputs, including datasets. The FracAtlas dataset is freely accessible under a CC-BY 4.0 license, allowing for widespread use in the scientific community, particularly among researchers and practitioners in medical imaging and related fields.
The data had created, cleaned, and managed by Iftekharul Abedeen, Md. Ashiqur Rahman, Fatema Zohra Prottyasha, Tasnim Ahmed, Tareque Mohmud Chowdhury & Swakkhar Shatabda. More details related to Data Collection & Annotation can be seen in ###Source Data section.

- **Repository:** https://figshare.com/articles/dataset/The_dataset/22363012
- **Paper:** https://www.nature.com/articles/s41597-023-02432-4

## Uses

<!-- Address questions around how the dataset is intended to be used. -->
The  "FracAtlas" dataset can be used to develop multiple machine learning or deep learning algorithms. For example: 
1. Developing a deep learning model to automatically detect fractures in radiographs.
2. Classifying the type of fractures (e.g., hairline, compound, transverse) using machine learning models
3. Implementing segmentation models to delineate bone structures from the surrounding tissues in the radiographs
4. Forecasting patients’ outcomes based on the characteristics of the fracture and other patient data
5. Developing models to identify anomalous patterns in the radiographs of bones


## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

### Original Dataset Schema
The original zip file contains 3 subfolders “images”, “Annotations”, “utilities” and a “dataset.csv” file. 
In the "image" folder, it contains 2 subfolders: "Fractured" and "Non-fractured", and each image is stored in corresponding folder in JPG format.
In the "Annotations" folder, it contains 4 subfolders: "COCO JSON", "PASCAL VOC", "VGG JSON", and "YOLO". Annotations are stored in their corresponding folders. More details can be read in ### Annotations section.
In the "utilities" folder, it contains several programming scripts that could be used to convert the raw files to a more readable format. None of them were used in this dataset loader.
The "dataset.csv" contains many basic variables for each image: <br />
- **image_id [string]:** The unique identifier for each radiograph image in the dataset. <br />
- **hand[int]:** A binary indicators (1 or 0) marking the presence of hand in the radiograph <br />
- **leg [int]:** A binary indicators (1 or 0) marking the presence of leg in the radiograph <br />
- **hip [int]:** A binary indicators (1 or 0) marking the presence of hip in the radiograph <br />
- **shoulder [int]:** A binary indicator (1 or 0) marking the shoulder in the radiograph <br />
- **mixed [int]:** A binary indicator of whether the image contains multiple body parts <br />
- **hardware [int]:** A binary indicator marking the presence of medical hardware (i.e. screws or plates) in the image <br />
- **multiscan [int]:** A binary indicator signifies whether the image is part of a set of multiple scans <br />
- **fractured [int]:** A binary indicator of whether there is a fracture present in the image <br />
- **fracture_count [int]:** The number of fractures present in the image <br />
- **frontal [int]:** A binary indicator denoting the front orientation of the radiograph <br />
- **lateral[int]:**  A binary indicator denoting the side orientation of the radiograph <br />
- **oblique [int]:** A binary indicator denoting denoting the angled orientation of the radiograph <br />

### Updated Dataset Schema
In this dataset loader, certain existed variables are extracted from the orginal "dataset.csv" and then modified into specific Huggingface feature class for clarity, for instance ClassLabel.
Other important variables are extracted from other downloaded files in the "FracAtlas" zip file to present a more systematic and clean FracAtlas dataset.
The full schema of the HuggingFace dataset loader is below:

- **image_id [string]:** The unique identifier for each radiograph image in the dataset. <br />
- **Image [image]:** A PIL image object denoting each X-ray image. This can be used to load the image file directly. <br />
- **hand[ClassLabel]:** A binary indicators (1 or 0) marking the presence of hand in the radiograph <br />
- **leg [ClassLabel]:** A binary indicators (1 or 0) marking the presence of leg in the radiograph <br />
- **hip [ClassLabel]:** A binary indicators (1 or 0) marking the presence of hip in the radiograph <br />
- **shoulder [ClassLabel]:** A binary indicator (1 or 0) marking the shoulder in the radiograph <br />
- **mixed [ClassLabel]:** A binary indicator of whether the image contains multiple body parts <br />
- **hardware [ClassLabel]:** A binary indicator marking the presence of medical hardware (i.e. screws or plates) in the image <br />
- **multiscan [ClassLabel]:** A binary indicator signifies whether the image is part of a set of multiple scans <br />
- **fractured [ClassLabel]:** A binary indicator of whether there is a fracture present in the image <br />
- **fracture_count [int]:** The number of fractures present in the image <br />
- **frontal [ClassLabel]:** A binary indicator (1 or 0) denoting the front orientation of the radiograph <br />
- **lateral[ClassLabel]:**  A binary indicator (1 or 0) denoting the side orientation of the radiograph <br />
- **oblique [ClassLabel]:** A binary indicator (1 or 0) denoting denoting the angled orientation of the radiograph <br />
- **localization_metadata [dict/Features]:** Metadata about the image localization, including 1) width(int), height (int), and depth (int) of the image
- **segmentation_metadata[dict/Features]:** Metadata about the segmentation, including the 1) segmentation(Sequence of Sequence of floats), 2) bounding box(Sequence of floats), and 3) area(float) covered by the segmentation. This can be None if no segmentation data is available

Also, we should note that even though the authors claim that annotations are provided only for images with fractures, it is worth-noting that some of the non-fracture images also have annotation data, and some of the fracture images do not. Therefore, to maximize the integrity of the data, both **Fractured** and **Segmentation_metadata** are kept for users. That is probably because annotations are done manlually and thus subject to errors, as the authors mentioned in the corresponding paper.
Furthermore, **hand**, **leg**, **hip**, and **shoulder** are not mutually exclusive, so they are stored as independent variables. 

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

The creation of the FracAtlas dataset was driven by the need for a comprehensive and specialized collection of medical images to train machine learning models for fracture detection. The dataset aims to address the gap in the availability of annotated musculoskeletal radiographs necessary for advancing AI-assisted diagnostic tools. The choices involved in its assembly, including the selection of specific types of radiographs and the detailed annotation process, were governed by the objective of providing a robust resource that can significantly improve the accuracy and efficiency of fracture diagnosis in the medical field. 

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
In the initial phase, a total of 14,068 X-Rays were collected. 
Due to privacy concerns, all the DICOM images were given an arbitrary image name and converted to JPG image format. 
This automatically got rid of all the sensitive information that was present in the metadata of DICOM images. 
These conversions were done using the proprietary software of the corresponding X-ray machines. 
The renamed DICOM images were stored in the hospital database separately for later study of general distribution. 
All the X-ray scans that have been collected are for general-purpose diagnosis. 
This means along with bone fracture scans there are also samples for chest diseases and abnormalities in the skull and spine region. 
In the collected data the number of bone fracture samples in the chest, skull and spine region was sparse. 
As a result, scans for the said parts were removed with the supervision of a medical officer. 
This left us with 4,083 scans from the hand, leg, hip and shoulder regions.

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

The FracAtlas dataset was accumulatively collected over 14,000 X-ray scans from several medical facilities across Bangladesh, with a substantial portion sourced from Lab-Aid Medical Center. 
Following collection, a meticulous data cleaning phase was undertaken to ensure the integrity and usability of the scans. 
Finally, the dataset was enhanced with detailed annotations. 
Ethical approval was secured, ensuring the confidentiality of patient data, and all participants provided informed consent. 
The collection process was designed to be non-intrusive to the standard diagnostic and treatment protocols of the involved hospitals.


### Annotations

<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
The dataset includes 4,083 images that have been manually annotated for bone fracture classification, localization, and segmentation with the help of 2 expert radiologists. 
Annotations have later been verified and merged by an orthopedist using the open-source labeling platform, makesense.ai. 
There are 4 types of annotations provided in the Annotations folder: 

  1. Common Objects in Context (COCO) JSON: It contains a JSON file, which includes corresponding annotations for fractured images (total 717 images). It includes segmentation, bbox, and area for each fractured image. This is mainly used for segmentation. Notice that the COCO annatation annotation is only for images that have fractures.
  2. PASCOL VOC: It contains xml files for each image. This is used for localization. For each xml file, it includes the height, width, depth, and segmented data for each image.
  3. VGG JSON: It contains a single JSON file, which includes annotations for fractrued images. 
  4. YOLO: It contains txt files for each image. This is used for localization.


#### Personal and Sensitive Information

<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->

All personally identifiable information in the gathered data has been removed, and theprocess was administered according to the Institutional Research Ethics Board of United International University.

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

While the FracAtlas dataset is particularly valuable for the development of computer-aided diagnosis systems, its potential limitations should be carefully considered. Firstly, the manual annotation process, is susceptible to human error, which may result in mislabeling. Such inaccuracies can impact the performance of machine learning models trained on this data. For example, the authors claim that annotations (segmentation, area, bounding box) are provided only for fracture images, some non-fractured images also have annotations. Conversely, some fracturd images miss corresponding annotations.  
It should be noted that to use the dataset correctly, one needs to have knowledge of medical and radiology fields to understand the results and make conclusions based on the dataset. It's also important to consider the possibility of labeling errors.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**APA:**

Abedeen, I., Rahman, M. A., Prottyasha, F. Z., Ahmed, T., Chowdhury, T. M., & Shatabda, S. (2023). FracAtlas: A Dataset for Fracture Classification, Localization and Segmentation of Musculoskeletal Radiographs. Scientific data, 10(1), 521. https://doi.org/10.1038/s41597-023-02432-4