Related Experiment Video
Updated: Nov 9, 2025
06:52
Stabilized Longitudinal In Vivo Cellular-Level Visualization of the Pancreas in a Murine Model with a Pancreatic Intravital Imaging Window
Published on: May 6, 2021
4.5K
Quality gaps in public pancreas imaging datasets: Implications & challenges for AI applications
Garima Suman1, Anurima Patra1, Panagiotis Korfiatis1
1Department of Radiology, Mayo Clinic, 200 First Street SW, Rochester, MN, 55905, USA.
Summary
Quality issues in public pancreas imaging datasets (PPIDs) limit AI research. Studies using these datasets often overlook these gaps, impacting results. We provide corrected segmentations to improve PPID utility.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Radiology
- Pancreatic Cancer Research
Background:
- Medical imaging datasets are crucial for AI development in diagnosing pancreatic cancer.
- Existing public pancreas imaging datasets (PPIDs) may contain quality gaps that affect research outcomes.
- Errors in experimental data can lead to profound inaccuracies in scientific findings.
Purpose of the Study:
- To characterize quality gaps in public pancreas imaging datasets (PPIDs).
- To evaluate the impact of these quality gaps on previously published studies.
- To provide enhanced post-hoc labels and segmentations for PPIDs to improve their utility.
Main Methods:
- Assessed PPIDs using the medical imaging data readiness (MIDaR) scale, evaluating metadata, image quality, acquisition phase, lesion etiology, confounders, and biases.
- Reviewed studies utilizing PPIDs to determine awareness and impact of quality gaps.
- Performed volumetric pancreatic adenocarcinoma (PDA) segmentations on CTs by a junior radiologist (R1) and reviewed by a senior radiologist (R3).
Main Results:
- Three PPIDs with 560 CTs and 6 MRIs were analyzed.
- While one dataset met high quality standards, others were rated lower due to incomplete annotations, limited metadata, and poor documentation.
- A significant proportion of CTs were unsuitable for AI due to factors like biliary stents, non-portal venous phase, or suboptimal image quality.
- None of the 25 studies using these PPIDs accounted for identified quality gaps.
- Radiologist R3 corrected 18% of PDA segmentations performed by R1.
Conclusions:
- Public pancreas imaging datasets exhibit substantial quality gaps and biases, with many CTs unsuitable for AI development.
- Published research using these datasets often fails to acknowledge or address these quality limitations.
- Post-hoc labels and segmentations are provided to enhance PPIDs, but collaborative efforts for larger, well-curated datasets are essential for advancing AI in clinical practice.
