Date: Tue, 05 Nov 1996 21:01:29 GMT Server: NCSA/1.5 Content-type: text/html Last-modified: Thu, 04 Jan 1996 22:45:24 GMT Content-length: 10933 Computerized Diagnosis Press Release

BREAST CANCER DIAGNOSIS1S VIA IMAGE ANALYSIS AND MACHINE LEARNING

W.H. Wolberg, W.N. Street, and O.L. Mangasarian

During the past six years, my colleagues and I at the University of Wisconsin, Madison have taken a mathematical technique originally developed for oil prospecting and modified it to create a computerized method of diagnosing cancer in breast tissue samples obtained by a technique called fine-needle aspiration (FNA). In the past, interpreting FNA samples has been rather subjective, but with our program we have been able to develop an accurate and objective system of FNA interpretation.

With this computer program, we are also able for the first time to calculate a mathematical probability that a sample is malignant, rather than having to use fuzzy terms such as atypical or suspicious. By sharing this probability data with patients, we involve them in the decision-making process to a greater degree than has previously been possible. In particular, instead of our having to make a recommendation based solely upon our own values, we can provide patients with the probability data and let them make up their own minds as to subsequent treatment.

While we don't expect that our computer program will be putting pathologists out of business any time soon, we do expect that it will prove useful as an objective backup for pathologists and assist them in improving their FNA interpretative skills.

THE ROLE OF FNA IN BREAST CANCER DIAGNOSIS

In the majority of all beast cancer cases, discovery of a mass -- feeling a lump in the breast -- represents the first sign that cancer is present. However, finding a breast mass does not necessarily signal that cancer is present: indeed, most breast masses are benign, not malignant.

Currently, the only definitive way to distinguish between benign and malignant breast masses is for a pathologist to examine a sample of breast tissue under a microscope. To obtain such tissue, surgeons have traditionally performed a breast biopsy, in which they take a patient to the operating room, administer anesthesia, cut open the breast, and remove a piece of tissue.

More recently, FNA has emerged as a less invasive, less painful, and less expensive alternative to surgical breast biopsy. As its name suggests, breast FNA involves inserting a small needle into the breast and suctioning cells into a syringe. The procedure takes only minutes and does not require anesthesia.

THE PROBLEM OF FNA SUBJECTIVITY

Until recently, however, interpretation of an FNA specimen has been a highly subjective process that, in the absence of firm objective diagnostic criteria, relies heavily on the training and experience of the persons examining the tissue. Although FNA has gained wide acceptance for the diagnosis of cancer in the thyroid and certain other organs, clinicians treating breast cancer have usually preferred to employ less subjective diagnostic procedures such as breast biopsy.

When looking under a microscope at pieces of breast biopsy tissue, pathologists can examine the intact structure of breast tissue and look for certain signs that provide definitive signs of malignancy. One such sign, for instance, is call invasiveness: if the pathologist sees a group of cells invading into normal tissue, that invasion by itself is a sufficient indicator that cancer is present.

The FNA technique, though, destroys the structure of breast tissues and thus forces pathologists to base their interpretation almost entirely on the appearance of the individual cells in the sample. Pathologists examining a breast FNA specimen need to evaluate a number of cell features, such as size, shape, and various nuclear characteristics. However, since no single one of these features, by itself, is able to yield an unequivocal diagnosis of malignancy, pathologists have had to subjectively weigh these various features in order to arrive at a diagnosis.

In instances when the various cell features point in the same direction, the diagnosis of malignant vs. benign is clear-cut. In other instances, though, when features point in different directions -- some suggesting malignant and others suggesting benign -- different observers might arrive at different verdicts, depending on how they choose to weigh the conflicting pieces of evidence. In the face of such potential for uncertainty and the consequences of a wrong diagnosis, clinicians have been reluctant to rely on FNA for diagnosing breast cancer.

PROSPECTING FOR CANCER

In 1987, during a chance encounter with Olvi L. Mangasarian, Ph.D., a professor in the Computer Sciences department at UW-Madison, I described to him the frustrations that I had encountered in trying to find a more objective way of interpreting FNA specimens. He realized that my problem was analogous to a linear programming problem he had studied some 20 years earlier while working in the oil industry -- the problem of where to drill for oil. Like evaluating FNA samples for cancer, deciding where to drill for oil involves finding a way to weigh a number of factors, such as geological features, no single one of which can definitively predict whether or not one will find oil.

To solve his oil prospecting problem, Dr. Mangasarian developed a method call multisurface pattern separation, which we have now adapted for use with FNA samples. In essence, the method begins with us mathematically modelling the cytological features, such as size or shape, that we need to evaluate in order to determine whether or not an FNA sample is malignant. Next we "train" the computer with data from two sets of FNA samples for which we already know the diagnosis -- one set of benign samples, and a second set of malignant samples. During this machine process, we iteratively build portions of a "fence" between the two data sets until we have completely fenced off the two data sets from each other.

Once we have trained the computer, we can then enter data from an unknown FNA sample. The computer determines which side of the "fence" the sample falls on and makes the appropriate diagnosis of malignant or benign. The computer also calculates a probability of malignancy which serves as a quantitative measure of the degree to which the computer is certain of its diagnosis.

IMAGE ANALYSIS

Using digital image analysis techniques developed by Mr. Nick Street, a Computer Science graduate student, we have now largely automated our computerized system for interpreting FNA samples. We look under a microscope to find suitable views of the sample and use a video camera attached to the microscope to record the images. The computer then digitizes the video pictures and stores the data in computer files.

To run an FNA analysis, the computer program reconstructs the digitized microscopic image and displays it on the computer screen. The computer operator simply uses a mouse to trace out rough outlines of the nuclei, and the computer does the rest. If desired, the operator can trace out more nuclei or remove nuclei that have already been traced out, and then return the analysis.

FINDINGS AND RESULTS

Our first goal has been to make the diagnosis of breast cancer from FNA samples more objective and accurate than has previously been possible. We have now succeeded in developing a highly accurate diagnostic system that requires only a modest level of expertise to operate. We now have a training set of 569 FNA samples and calculate that the system will correctly diagnose breast FNA samples 97% of the time. In practice, the system has correctly diagnosed the last 92 samples that we have tested.

Our second goal has been to develop an objective method of determining prognosis for those patients whom our diagnostic program identifies as having breast cancer, that is, to accurately predict the likelihood that their cancer will recur. In regards to this goal, we are still at a relatively early stage of development, as we only have 187 samples in our training set. Although we are encouraged by our results to date, we still need to add more samples to our training set and to improve the program's accuracy.

One of the key findings from our studies has been the discovery that with our computer program, nuclear features (i.e. nuclear "grade") are more accurate in determining prognosis than are determinations of prognosis based on the traditional measures of tumor size and axillary node status. Our research may thus have a significant impact on the current practice of surgeon performing axillary lymph node dissection at the time of mastectomy for prognostic purposes.

Presently, determination of whether breast cancer has spread into a patient's axillary lymph nodes is considered to be an important factor in establishing prognosis: patients with tumor spread in axillary lymph nodes (e.g. node positive) have a higher likelihood of tumor recurrence than those without tumor spread into axillary lymph nodes (e.g. node negative). Patients with node positive tumors tend also to have earlier recurrences and, in general, do worse than patients with node negative tumors.

Data from our research, however, have indicated that a patient's nodal status (node positive vs. node negative) provides no new additional information about prognosis beyond that which is already available from an FNA biopsy. When we use FNA data alone to run our prognostic program, the results are just as accurate as when we run the program using FNA data plus information on nodal status and tumor size. That is, we have found that analysis of a patient's FNA sample alone provides all the information about prognosis that we need, and thus that performing an axillary lymph node dissection solely for prognostic purposes is unnecessary. Should studies at other institutions confirm to our findings, many breast cancer patients in the future could be spared the necessity of undergoing axillary lymph node biopsy.

FUTURE DIRECTIONS

We are now seeking to collaborate with other institutions to validate our computer program using their FNA samples. We have not patented our program and are willing to provide copies of our program to those researchers who sign collaborative agreements with us. We recently analyzed 19 breast FNA slides from UCLA and were gratified to find that we correctly diagnosed them all.

We are are also interested in expanding our computerized interpretative method to other organs besides breasts. We have had good early results with thyroid cancer and also believe our approach could prove particularly useful for evaluating lymphomas.


Last modified: Thu Jan 4 16:45:23 1996 by Nick Street
street@cs.wisc.edu