User Stories

Decision Support Technique Developed in MATLAB Improves the Accuracy of Medical Diagnoses

Challenge

Analyzing and interpreting the huge quantities of available data is a taxing problem for today's diagnostician. Using traditional spreadsheets and tables, it is difficult to identify the trends, anomalies, and other significant features in the data that are essential to correct diagnosis.

The University of Sheffield research team, led by Dr. Rob Harrison of the Department of Automatic Control and Systems Engineering and Dr. Simon Cross of the Department of Pathology, set out to create a more effective diagnostic tool. "The human brain is far better equipped to interpret visual images than rows of numbers," says Cross, a breast cancer specialist. "Our aim was to devise a system that could convert statistics into images that would be easier for researchers and diagnosticians to analyze."

Neural networks were ideal for this type of work, explains Harrison: "They can cope with noisy data and, unlike some other systems, they begin by assuming nothing, so they can sometimes identify anomalies that might defeat other systems. They provide a practical, day-to-day tool for reaching the conditional mean without becoming too bogged down in theory."

"MATLAB contains extremely sophisticated math and engineering functions, and it is easy to add your own. The result is a tool that allows engineers to produce designs in the field and gain a quick, reliable understanding of what they are observing."

Dr. Rob Harrison
University of Sheffield
Department of Automatic Control and Systems Engineering

Solution

Dr. Harrison, who began using MATLAB in 1986 and has "never coded in anything else since," recommended MATLAB for this project. Standard neural networks deal with multi-dimensional data, which is difficult for the human brain to assimilate. MATLAB reduces this data to 3-D graphs and plots with overlays that can be readily understood even by inexperienced observers. An additional benefit, says Harrison, is that "MATLAB is effortless to learn and use and lets students get to grips with sophisticated ideas within a very short time." Dr. Cross, who had never encountered MATLAB before, learned to write short functional MATLAB programs and use MATLAB to manipulate the color graphic output virtually overnight.

The technique they developed uses a growing cell structure (GCS) network, an unsupervised self-learning artificial neural network. A graphical user interface developed in MATLAB allows clinicians to quickly start using this method. GCS sorts the data into types (similar cases), which become nodes on the MATLAB graph. No reference is made to the outcome while the data is being sorted. Only after the system has attained its final size is each case linked to its known outcome. Statistical methods are used to place new cases correctly on the graph. The frequencies of cases with each outcome can predict the outcome of any new case placed at that node on the graph.

To test the technique, the researchers selected 692 breast lump specimens, noting the patient's age and ten other pertinent variables. They used either an open biopsy or information from mammograms, combined with the absence of further malignancy, to confirm the final outcome of benign or malignant cancer. The data was randomly partitioned into a training set of 462 and a case-test set of 230. Using conventional techniques, they derived a logistic equation from the training set which they then applied to the case-test set.

Next, working in MATLAB, they applied the GCS methodology to the same data. MATLAB produced a network that contained areas of high-certainty positive and negative outcomes, with a narrow area of transition between them. This result compared very favorably with that obtained using the conventional logistic regression technique (96% as opposed to 98%).

"The new technique does not replace human interpretation of data, but provides a powerful decision support mechanism to help validate the results," says Dr Cross. "It is the most useful technique we have had for years."

Results

Products Used