Hui Luo, PhD, Carestream Health Inc.; Jacquelyn Ellinwood; David Foos; Eliot Siegel, MD, FSIIM; Bruce Reiner, MD

The increasing pressure to improve quality and lower costs is forcing radiology departments to invest heavily on new technologies, such as PACS and RIS. Currently, these new technologies are primarily used for collection and storage of patients’ data to assist in image diagnosis. Once the diagnosis is accomplished, the data “sleep” in the databases and are rarely retrieved for other purposes.

In addition to providing diagnosis information, the image database, as a whole, also contains other “hidden” information that, although not directly associated with diagnosis for a particular patient, may have value related to overall image quality and performance of a department’s operations, including information of value to management, medical education and staff training, and research. Today, the administrators and managers of radiology departments are frequently asked to evaluate their department’s operations, and identify opportunities for improvement. Unfortunately, most radiology department operations are currently estimated from the limited, outdated, inaccurate and sometimes unreliable data, which makes it difficult to represent the true performance of the department. Data mining image databases provides significant benefits for obtaining up-to-date, accurate and valuable information about a department’s operations, enabling data-driven performance improvement, and enhancing the quality of health care delivery.

The goal of this work is to develop a user-friendly data mining tool for the administrative staff in radiology departments to seek out valuable information from their vast databases, and to assist them in developing effective strategies for quality improvement, staff training, and operation management.

Methods:

A data mining tool for discovering image quality information from image database is developed, and it includes the following five components:

1) Data source: communicates with and provides access to data that is stored in different databases, such as PACS, RIS, CR or DR capture equipments.

2) Data processing engine: receives and processes the data from the data source based on instructions from users, and places the results into the processing database. Based on user interest and information requests, different processing methods are performed. For image quality information, the data processing engine detects image defects using a set of image defect detection algorithms, such as clipped anatomy detection, motion blur detection, and under or over exposure detection, as well as providing a retake rate calculation.

3) Data mining engine: extracts the information and/or discovers the hidden patterns or relationships in the processed data per users’ instructions and places the results in the data mining database. The information, pattern, or relationship provided by the data mining engine relates to what the user seeks. The information may be previously unknown, and has the potential of being very useful for users.

4) User instruction engine: provides a user-friendly interface to communicate with users, and outputs the requested information in the way that users desire.

Results:

A database of chest radiographic images was developed over a one-year period, based on sequential case collections at two high-volume imaging centers. The database was populated with all QA-accepted images (i.e., images determined by the radiographic technologist during the visual QA process as having acceptable diagnostic quality, and released to PACS). The proposed data mining tool processes all images in the database. If an image is evaluated by the data processing engine as having one or more defects, the associated image information along with the defect results, will be stored in the data processing database. The data mining engine extracts all defective images, analyzes their defect types and associated information, and outputs the discovered information in the user desired output format.

The initial version of the data mining tool focuses on two major issues. One is the summary of image quality at a site. The other is the performance profile of technologists. The summary of image quality during a period of time within the radiology department gives supervisors or administrative staff a systematic view of image quality in the department, helps them get insight into the occurred or potential problems, and enables them to develop an effective solution or be proactive in solving problems.

The technologist performance is another important component of the image quality assurance. A performance profile of each technologist can help a supervisor or administrative staff members to pinpoint the skill strength or weakness of a technologist and plan an effective educational and training plan for the technologist, if needed.

Data mining in medical databases is a fairly new topic. Recently, few data mining methods have been developed in medical databases. Most of the methods study the patient medical reports to aid in diagnosis of an individual patient. To our knowledge, no attention has previously been given to discovering the health care performance information from large-scale, in-department databases. In addition, there is no prior work on performing data mining in medical image databases, mainly because it is difficult to extract information from image data. The proposed data mining system provides a novel solution for discovering the image quality information from large image databases. The initial experiment results are promising.