TL;DR
This paper introduces MedImg, a comprehensive, open-access online database that integrates over 1.6 million medical images from 103 diverse public datasets. MedImg aims to address the critical challenge of data scarcity in medical image analysis by providing a unified platform for researchers to access, browse, and download a wide range of imaging modalities and organ types, thereby facilitating the development of more generalized and robust deep learning models.
What Problem Does This Paper Address?
The rapid advancements in deep learning have shown immense promise in medical image analysis, often achieving or even exceeding human performance. However, a significant hurdle to translating these powerful models into clinical practice is the lack of large-scale, well-characterized datasets. This scarcity makes it difficult to train, validate, and generalize deep learning models effectively. While various medical image databases exist, they often focus on a single organ, disease, or imaging modality, hindering the development of models that can handle the complexity and diversity of real-world medical imaging. MedImg directly tackles this by creating a centralized, multi-modal, and multi-organ repository.
Methods
The core of this paper is the creation and presentation of the MedImg database. The authors systematically collected medical image datasets from public sources, specifically targeting those with public licenses. They employed keywords like "medical image" to identify relevant datasets. A rigorous inclusion criterion was applied, resulting in the selection of 103 datasets out of an initial pool of 492. These datasets were then integrated and hierarchically organized within the MedImg online platform. The database is built on an Apache Tomcat server, with a front-end developed using HTML5 and CSS3, interactive functions powered by JQuery, and a back-end managed by a Python Django framework. The platform is designed for intuitive browsing, querying, and downloading of images, accessible across multiple platforms without requiring registration.
Datasets
MedImg is built upon 103 publicly available medical image datasets, totaling over 1.6 million images. These datasets cover:
- 14 Modalities: Including X-ray (XR), Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Optical Coherence Tomography (OCT), Ultrasound, Endoscopy (ES), Histopathology Slides (HS), and others.
- 9 Organs: Primarily focusing on Lung, Brain, Eye, Heart, Skin, Kidney, Liver, Prostate, and Knee.
- Diverse Tasks: Datasets are categorized by their primary deep learning task, with Classification (53%) and Segmentation (33%) being the most prevalent. Other tasks include Detection, Regression, Registration, Reconstruction, and Tracking.
- Data Sources: The datasets are mainly derived from websites hosting AI-related competitions, such as Grand Challenge and Kaggle.
- Annotation: 79 of the datasets include well-annotated labels, which are crucial for supervised learning tasks.
- Timeframe: Datasets were collected between 2007 and 2022, with a significant portion established after 2015, reflecting the surge in deep learning research.
MedImg itself is not a new benchmark dataset but rather an integration of existing ones, making them more accessible and usable for research.
Key Insights and Ideas
- The critical need for data aggregation: The paper strongly emphasizes that the advancement of deep learning in medical imaging is fundamentally bottlenecked by data availability. MedImg's contribution lies in its comprehensive aggregation, making it easier for researchers to find and utilize diverse data.
- Diversity is key for generalization: By integrating images across various modalities, organs, and even geographic areas (implied by public sources), MedImg aims to foster the development of models that are more robust and generalizable to unseen data.
- Accessibility drives research: The open-access and intuitive nature of MedImg, with no registration required, is designed to lower the barrier to entry for researchers, encouraging wider participation and collaboration.
- The evolving landscape of medical imaging data: The paper highlights the increasing availability of large datasets, particularly in recent years, driven by the success of deep learning itself and the growing recognition of the importance of data sharing.
Key Results
The primary "result" of this paper is the establishment and description of the MedImg database. The database contains:
- 103 Datasets
- 1,622,956 Images
- 14 Modalities
- 9 Organs
The paper details the distribution of these datasets across modalities (XR being the most frequent at 34%), organs (Lung dominating at 46%), and tasks (Classification at 53%, Segmentation at 33%). It also shows the distribution of dataset sizes, with a notable number of datasets containing over 5,000 or 10,000 images. The database is freely accessible at https://www.cuilab.cn/medimg/.
Limitations and Future Work
The authors acknowledge two primary limitations:
- Lack of Clinical Information: The current version of MedImg is short on detailed clinical information that would be beneficial for AI-based diagnosis.
- Manual Curation and Normalization: The datasets are not uniformly curated or normalized with a standard protocol, which can complicate the preprocessing step for users.
Future work will focus on:
- Incorporating New Datasets: Continuously updating the database with new medical image datasets.
- Enhancing Data Quality: Manually curating and normalizing all datasets with a standard protocol to simplify preprocessing for users.
- Adding Clinical Information: Potentially integrating more clinical data to enrich the datasets for diagnostic applications.
Summary updated