Navigation  without Java Scripts

Prolog Assists in Brain Lesions Diagnosis:

An MR and CT Features-Based Expert System

by Pasteur Rasuli, Firooz Rasouli, Ali Oskouie, Tannaz Rasouli, William F.Morrish

 

MRI-WIZ software was developed to determine the technical feasibility of an expert system for differential diagnosis in neuroradiology. This frame based expert system, developed in Visual Prolog, stores the Magnetic Resonance (MR) and Computerized Tomography (CT) imaging characteristics of over 100 known brain disorders in object-like entities. The frames, organized in a hierarchical relationship, have the highest frames containing information that applies to all lower frames. A decision tree menu provides a user-friendly interface that navigates through the network based upon lesion features as depicted on MR and CT images. Once the appropriate node on the tree menu is chosen, the user responds to CT and MR related questions - to obtain information about the lesion. The exact location of the brain lesion, as seen on the MR and CT images, is determined by the user on the color coded, hot-linked cross sectional diagrams of the brain. Program execution follows a consultation paradigm using a dynamic database and provides a differential diagnosis based on MR findings alone. Adding CT-related information may further refine the diagnosis. The program utilizes extensive on-line hypertext-based help information about the lesions with actual CT and MR images.

Introduction

Expert systems, using knowledge and inference techniques, mimic human reasoning and decision-making to solve problems. Production and frame-based systems are two basic types of expert systems1. In the production system, the knowledge is stored as a large body of facts2, along with rules expressing relationship between these facts. These programs base their decisions on rules and facts typically represented in the form of if-then rules. Each rule relates information that is known or could be known about a particular domain to something that can be inferred from that knowledge. By combining the applicable rules, the expert system can reach a conclusion or make a diagnosis. The frame-based system, an efficient and powerful form of data structure, represents contextual information about objects, concepts, or events. The frames, organized in a hierarchical structure, allow lower order frames to inherit attributes from higher order frames, with the highest frame containing information that applies to all other frames.

Expert systems have been assisting clinicians in areas such as infectious disease therapy for many years. While radiological technology relies heavily on computers, it has been slow to develop an expert system to aid in diagnosis. In 1992 we developed BONEX, a menu-driven DOS-based expert system for the radiological diagnosis of bone tumors using PDC Prolog. BONEX, a rule-based expert system, leads the user through a menu of questions, generates a hard copy report and a list of diagnosis with an estimate of the likelihood of each.

The success of BONEX was very well received at the 1992 annual meeting of the Radiological Society of North America (RSNA)3 which led us to consider the application of an expert system for the diagnosis of brain lesions. The objective of developing a program that could be accessed directly from MR imaging or CT console required the development of a Windows-based program with a programming environment that had previously been unavailable.

System Methodology and Architecture

In neuroradiology, MR imaging is the most prevalent imaging modality for the diagnosis of cerebral lesions. Useful adjunctive information is often obtained with CT and when integrated with the patient's age and clinical symptoms, the radiologist suggests the diagnosis or develops a short list of differential diagnoses4. The diagnostic process has limitations, however, resulting from human error caused by a lack of knowledge, a nonsystematic approach, or even forgetfulness5. An attractive solution to this dilemma is the use of artificial intelligence based technology, such as an expert system, that draws conclusions based on facts stored in a knowledge base. However, even the best expert systems have limitations when compared with the human expert, who is often able to apply intuition and common sense to solve problems when no formal solutions or analogy exists.

To determine the technical feasibility of an expert system in diagnosing brain lesions, the MRI-WIZ program was developed. Prolog was selected as the language of choice for implementing the program. Prolog programs, a collection of facts and rules about a particular knowledge domain, are unique in their ability to infer facts and conclusions from other facts. The program is really a database, with Prolog the very powerful query language. The database characteristics make the language ideal for problems that are unstructured and for which the procedures to solve it are unknown. Both production and frame-based expert systems can be built with Prolog. MRI-WIZ uses a frame-based expert system to store MR and CT imaging characteristics of over 100 known brain disorders4. The prototype version of the MRI-WIZ for Windows was developed using Visual Prolog based on its fully visual programming capabilities and cross-platform portability. The other attractive features of Visual Prolog include tools for making Windows help files and facilities for creating dynamic web pages. The program was field tested on proven cases at Ottawa General Hospital (Ontario, Canada) as the prototype program was being refined. The system was also showcased at the Department of the Future exhibit at the 1995 scientific assembly and annual meeting of the RSNA6. In this article, steps that were taken to refine and expand the program after the field testing and showcasing are discussed.

Program Refinement and Expansion

Based on discussions with several neuroradiologists during and after the 1995 exhibit, and the intensive verification and validation of the program on proven cases at Ottawa General Hospital, a 400-bed teaching hospital at The University of Ottawa, the program was modified to satisfy the end users' requirements. The modifications included cosmetic and conceptual changes to the program as well as knowledge enhancement of the database to improve the systems credibility. The changes include:

Location Menu - A graphical method of location identification was implemented providing a list of possible locations. The diagram (Figure 1), divided into eight segments, illustrates eight cross sectional views of the brain. The different regions of the brain are color coded and linked to a special database. When a region is selected via the mouse, the program displays the name of the region and asks for confirmation. Main Tree-Menu - A decision tree menu provides a user-friendly interface that navigates through the response network. Since the program provides a differential diagnosis based on MR imaging alone and is further refined by CT-related information, there were originally separate nodes for MR and CT options on each branch. As a result, the opening tree-menu looked complex and confusing. In the new version, the MR and CT nodes are combined (Figure 2), and a Not Known option was added to the response boxes enabling the radiologist to use the program without the CT information.

 fig1.gif (42817 bytes)

Figure 1: Location menu for identifying the site of the lesion

 fig2.gif (9010 bytes)

Figure 2: MRI-WIZ Opening Menu- Decision Tree

 Diagnosis Report - In the new version, in addition to the list of diagnoses, the user response to various questions is also summarized in an abbreviated form on top of the diagnosis report (Figure 3). If an abbreviation is in question, a "glossary" furnishes the user with abbreviations and meanings with the click of the mouse (Figure 4).

 fig3.gif (12033 bytes)

Figure 3: The diagnosis report-list of diagnoses and the user's response to various questions

 fig4.gif (6608 bytes)

Figure 4: Glossary of the abbreviated terms

 Online Help - One of the main features of the program is an extensive online hypertext-based help system containing information on all the lesions in the knowledge base. Users can refer to the comprehensive help menu for clarification purposes, or for a "hands-on" approach to identify brain lesions. The general description of a lesion under the help topic consists of: population at risk, prognosis, and differential diagnosis. Topics are accompanied by actual MR and CT images (Figure 5).

 fig5.gif (41941 bytes)

Figure 5: Help Menu-Portions of the online text information about Meningioma

 Modular Design - MRI-WIZ is modular with different aspects of the program and knowledge base in separate modules. This proves to be very helpful during the program modification and expansion process. The approach is also very useful during the program validation process since it allows for use of different versions of the knowledge base with minimal effort.

 Using MRI-WIZ

MRI-WIZ opens with a tree menu displaying various nodes; each node represents a possible signal intensity of a lesion. Before selecting the signal intensity the user identifies the location of the lesion by using the mouse on a color-coded diagram with eight cross sectional views of the brain.

Once the location is identified, the user is presented with a series of possibilities in the form of a decision tree. Begining with a broad array of options, the program gradually becomes more specific as each affirmed characteristic leads to a new series of options.

First, the user must identify whether the signal intensity of the lesion on T1-weighted images is Dark, Bright, or Isointense (relative to the neighboring normal brain tissue). If chosen, a fourth option, Other, allows the user to select from: Fluid Level (bi-intensity), Mixed intensity, Cystic intensity, Signal Void, or Hypointense Rim on T2. After describing the T1-weighted image, the user is presented with similar options for the T2-weighted image, Dark or Bright.

Next, several questions appear about the lesion.These questions can be yes-or-no, or contain options (Figure 6). In the CT related questions, a Not Known alternative is available to enable the user to pursue diagnosis in the absence of the CT information on the basis of MR information only. Required information includes data about the contrast enhancement, calcification, CT density, perilesional edema, hemorrhage, multiplicity, size, and hydrocephalus or ventriculomegaly. At any point throughout the procedure, the user may opt to cancel the search.

After completing the survey, the questions and user-provided answers are displayed together in abbreviated form. Beneath the synopsis of responses, the possible diagnoses are listed. The program also specifies if the lesion occurs in adult or pediatric populations, or in both.

 fig6.gif (4953 bytes)

Figure 6: Response box with multiple options

 In addition to the step-by-step, diagnosis application of the program, users may utilize MRI-WIZ as a reference. The Help menu not only offers operational support, but also extends detailed descriptions of all lesions in the knowledge base, classified by location of the lesion, category of the lesion, and age of the patient. The online help can also be accessed from the alphabetical list or through a simple search command.

 Conclusion

The introduction of computers to radiological imaging revolutionized radiology in the 60's and 70's with the development of Computer Assisted Tomography (CAT scanning), which later was extended to sonographic imaging, magnetic resonance imaging and digital subtraction angiography. The role of computers in radiological imaging, until recently, however, has mainly been confined to processing complex matrix data of radiological densities to produce gray scale digital images. The introduction of expert systems promises another major advancement brought to radiology by computers, this time in the area of diagnosis. Although it will be possible to train computers to scan images and make diagnosis independently, it is more accurate and, ultimately, safer to have the radiologist view images and feed the information to the computer. In this form of application, the radiologist also acts as the final referee in accepting or rejecting the diagnostic possibilities rendered by the computer. As we have shown, expert systems can play a vital role in advancing the field of radiological diagnosis.

 BIO info

Pasteur Rasuli, MD, FRCP(C), Associate Professor, Department of Radiology, The Ottawa Hospital-General Campus,University of Ottawa (Ottawa, Ontario). He can be reached at (613)-737-8098

Firooz Rasouli, PhD, currently at the Research and Development & Engineering Center, Philip Morris USA (Richmond, Virginia). He can be reached at rasoulif @ cs.com

Aki Oskouie, PhD, Department of Chemical & Environmental Engineering, Illinois Institute of Technology (Chicago, Illinois)

Tannaz Rasouli, currently studying at the Department of Cognetive Science, Johns Hopkins University (Baltimore, Maryland)

William F. Morrish, MD, FRCP(C), Neuroradiologist, Department of Diagnostic Imaging, Foothills Hospital (Calgary, Alberta)

 References

1. Townsend C. Introduction to Expert Systems. In:Townsend C.,ed. Introduction to Turbo Prolog. San Francisco, CA: SYBEX, 1986:213-219

2. Theft L. Overview of Knowledge Based Systems. In Teft L, ed. Programming in Turbo Prolog with an Introduction to Knowledge Based Systems. Englewood Cliffs, NJ: Prentice-Hall, 1989; 1-21

3. Rasuli P, Rasouli F. Diagnosis of Bone Tumors with a Personal Computer (abstr.). Radiology 1992; (185(P):413

4. Rasuli, P, Rasouli F, Hammond, I, Amiri, F An Artificial Intelligence Program for the Radiologic Diagnosis of Brain lesions. RadioGraphics 1996, 16,5,1207-1213

5. Lodwick GS. "Radiologic concepts. In: Hodes PJ, ed. The Bones and Joints: an Atlas of Tumor Radiology. Chicago, IL:Year Book Medical, 1971;1-82

6. Rasuli, P, Rasouli F, Amiri, F Department of the Future Exhibit, RSNA Scientific Assembly and Annual Meeting, Chicago, IL, Nov 1995