LDCTiqa_png: Low-dose CT Image Quality Assessment Dataset

Dataset Description

This dataset, MikaJesse/LDCTiqa_png, contains 1000 PNG images of low-dose computed tomography (CT) slices, each paired with an expert-provided perceptual image quality score. The data originates from the Low-dose Computed Tomography Perceptual Image Quality Assessment (LDCT-IQAC) 2023 Grand Challenge.

This resource is designed to facilitate research and development in No-Reference Image Quality Assessment (NR-IQA) for medical imaging, particularly for low-dose CT scans.

Background and Motivation

Image quality assessment (IQA) is extremely important in CT imaging, as it facilitates radiation dose optimization and the development of novel algorithms like image restoration. While widely used metrics like PSNR and SSIM exist, their correlation with radiologists’ opinions is often insufficient, and they require pristine reference images which are unavailable in real clinical settings.

This dataset was created to address the lack of open-source datasets for CT IQA and to provide a standardized benchmark. The images in this challenge specifically deal with a combination of two major artifacts common in low-dose CT: sparse view streak artifacts and noise. The goal is to develop and evaluate NR-IQA models that correlate well with the perceptual scores from expert radiologists.

Dataset Contents

• Images: 1000 CT slices, originally from the challenge's training set, converted to PNG format for ease of use.
• Scores: A JSON file mapping each image filename to its corresponding quality score.

Image Scoring Criteria

The image quality scores were obtained from five experienced radiologists using abdominal soft-tissue windows (width/level: 350/40). The final human perceptual score for each image is the average of the scores assigned by the five radiologists. The scores range from 0 (very poor) to 4 (excellent).

How to Use

The dataset can be easily loaded using the Hugging Face datasets library.

from datasets import load_dataset
from PIL import Image

# Load the dataset
dataset = load_dataset("MikaJesse/LDCTiqa_png")

# Access an example
example = dataset['train']

image: Image.Image = example['image']
score: float = example['score']
filename: str = example['image'].filename # You can get the filename like this

print(f"Image Filename: {os.path.basename(filename)}")
print(f"Image Size: {image.size}")
print(f"Expert Score: {score}")

# To display the image (e.g., in a Jupyter notebook)
# display(image)

Data Fields

Each example in the dataset consists of two main fields:

• image: A PIL.Image.Image object containing the 2D CT slice.
• score: A float32 value representing the average perceptual quality score from radiologists (0-4).

Data Splits

The dataset contains a single split, train, which includes all 1000 images and their corresponding scores.

Source Data

This dataset is a reformatted version of the data provided for the LDCT-IQAC 2023 Grand Challenge.

• Original Challenge Website: https://ldctiqac2023.grand-challenge.org/

Citation Information

If you use this dataset in your research, please be sure to cite the original paper that introduced the challenge and dataset:

@article{lee2025low,
  title={Low-dose computed tomography perceptual image quality assessment},
  author={Lee, Wonkyeong and Wagner, Fabian and Galdran, Adrian and Shi, Yongyi and Xia, Wenjun and Wang, Ge and Mou, Xuanqin and Ahamed, Md Atik and Imran, Abdullah Al Zubaer and Oh, Ji Eun and others},
  journal={Medical Image Analysis},
  volume={99},
  pages={103343},
  year={2025},
  publisher={Elsevier}
}

Licensing

The license for this dataset is not explicitly specified on the Hugging Face Hub. Please refer to the original data source at the LDCT-IQAC 2023 Grand Challenge website to determine the usage rights and licensing terms. It is the user's responsibility to ensure compliance with the original license.