http:^^www.cs.wisc.edu^~street^press_rel.html
来自「This data set contains WWW-pages collect」· HTML 代码 · 共 249 行
HTML
249 行
Date: Tue, 05 Nov 1996 21:01:29 GMTServer: NCSA/1.5Content-type: text/htmlLast-modified: Thu, 04 Jan 1996 22:45:24 GMTContent-length: 10933<HTML><HEAD><TITLE>Computerized Diagnosis Press Release</TITLE></HEAD><BODY><H1>BREAST CANCER DIAGNOSIS1S VIA IMAGE ANALYSIS AND MACHINE LEARNING</H1><H2>W.H. Wolberg, W.N. Street, and O.L. Mangasarian</H2><P>During the past six years, my colleagues and I at theUniversity of Wisconsin, Madison have taken a mathematicaltechnique originally developed for oil prospecting and modified itto create a computerized method of diagnosing cancer in breasttissue samples obtained by a technique called fine-needleaspiration (FNA). In the past, interpreting FNA samples has beenrather subjective, but with our program we have been able todevelop an accurate and objective system of FNA interpretation.<P>With this computer program, we are also able for the firsttime to calculate a mathematical probability that a sample ismalignant, rather than having to use fuzzy terms such as atypicalor suspicious. By sharing this probability data with patients, weinvolve them in the decision-making process to a greater degreethan has previously been possible. In particular, instead of ourhaving to make a recommendation based solely upon our own values,we can provide patients with the probability data and let them makeup their own minds as to subsequent treatment.<P>While we don't expect that our computer program will beputting pathologists out of business any time soon, we do expectthat it will prove useful as an objective backup for pathologistsand assist them in improving their FNA interpretative skills.<H2>THE ROLE OF FNA IN BREAST CANCER DIAGNOSIS</H2>In the majority of all beast cancer cases, discovery of a mass-- feeling a lump in the breast -- represents the first sign thatcancer is present. However, finding a breast mass does notnecessarily signal that cancer is present: indeed, most breastmasses are benign, not malignant.<P>Currently, the only definitive way to distinguish betweenbenign and malignant breast masses is for a pathologist to examinea sample of breast tissue under a microscope. To obtain suchtissue, surgeons have traditionally performed a breast biopsy, inwhich they take a patient to the operating room, administeranesthesia, cut open the breast, and remove a piece of tissue.<P>More recently, FNA has emerged as a less invasive, lesspainful, and less expensive alternative to surgical breast biopsy.As its name suggests, breast FNA involves inserting a small needleinto the breast and suctioning cells into a syringe. The proceduretakes only minutes and does not require anesthesia.<H2>THE PROBLEM OF FNA SUBJECTIVITY</H2>Until recently, however, interpretation of an FNA specimen hasbeen a highly subjective process that, in the absence of firmobjective diagnostic criteria, relies heavily on the training andexperience of the persons examining the tissue. Although FNA hasgained wide acceptance for the diagnosis of cancer in the thyroidand certain other organs, clinicians treating breast cancer haveusually preferred to employ less subjective diagnostic proceduressuch as breast biopsy.<P>When looking under a microscope at pieces of breast biopsytissue, pathologists can examine the intact structure of breasttissue and look for certain signs that provide definitive signs ofmalignancy. One such sign, for instance, is call invasiveness: ifthe pathologist sees a group of cells invading into normal tissue,that invasion by itself is a sufficient indicator that cancer ispresent.<P>The FNA technique, though, destroys the structure of breasttissues and thus forces pathologists to base their interpretationalmost entirely on the appearance of the individual cells in thesample. Pathologists examining a breast FNA specimen need toevaluate a number of cell features, such as size, shape, andvarious nuclear characteristics. However, since no single one ofthese features, by itself, is able to yield an unequivocaldiagnosis of malignancy, pathologists have had to subjectivelyweigh these various features in order to arrive at a diagnosis.<P>In instances when the various cell features point in the samedirection, the diagnosis of malignant vs. benign is clear-cut. Inother instances, though, when features point in differentdirections -- some suggesting malignant and others suggestingbenign -- different observers might arrive at different verdicts,depending on how they choose to weigh the conflicting pieces ofevidence. In the face of such potential for uncertainty and theconsequences of a wrong diagnosis, clinicians have been reluctantto rely on FNA for diagnosing breast cancer.<H2>PROSPECTING FOR CANCER</H2>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 encounteredin trying to find a more objective way of interpreting FNAspecimens. He realized that my problem was analogous to a linearprogramming problem he had studied some 20 years earlier whileworking in the oil industry -- the problem of where to drill foroil. Like evaluating FNA samples for cancer, deciding where todrill for oil involves finding a way to weigh a number of factors,such as geological features, no single one of which candefinitively predict whether or not one will find oil.<P>To solve his oil prospecting problem, Dr. Mangasariandeveloped a method call multisurface pattern separation, which wehave now adapted for use with FNA samples. In essence, the methodbegins with us mathematically modelling the cytological features,such as size or shape, that we need to evaluate in order todetermine whether or not an FNA sample is malignant. Next we"train" the computer with data from two sets of FNA samples forwhich we already know the diagnosis -- one set of benign samples,and a second set of malignant samples. During this machineprocess, we iteratively build portions of a "fence" between the twodata sets until we have completely fenced off the two data setsfrom each other.<P>Once we have trained the computer, we can then enter data froman unknown FNA sample. The computer determines which side of the"fence" the sample falls on and makes the appropriate diagnosis ofmalignant or benign. The computer also calculates a probability ofmalignancy which serves as a quantitative measure of the degree towhich the computer is certain of its diagnosis.<H2>IMAGE ANALYSIS</H2>Using digital image analysis techniques developed by Mr. NickStreet, a Computer Science graduate student, we have now largelyautomated our computerized system for interpreting FNA samples. Welook under a microscope to find suitable views of the sample anduse a video camera attached to the microscope to record the images.The computer then digitizes the video pictures and stores the datain computer files.<P>To run an FNA analysis, the computer program reconstructs thedigitized microscopic image and displays it on the computer screen.The computer operator simply uses a mouse to trace out roughoutlines of the nuclei, and the computer does the rest. Ifdesired, the operator can trace out more nuclei or remove nucleithat have already been traced out, and then return the analysis.<H2>FINDINGS AND RESULTS</H2>Our first goal has been to make the diagnosis of breast cancerfrom FNA samples more objective and accurate than has previouslybeen possible. We have now succeeded in developing a highlyaccurate diagnostic system that requires only a modest level ofexpertise to operate. We now have a training set of 569 FNAsamples and calculate that the system will correctly diagnosebreast FNA samples 97% of the time. In practice, the system hascorrectly diagnosed the last 92 samples that we have tested.<P>Our second goal has been to develop an objective method ofdetermining prognosis for those patients whom our diagnosticprogram identifies as having breast cancer, that is, to accuratelypredict the likelihood that their cancer will recur. In regards tothis goal, we are still at a relatively early stage of development,as we only have 187 samples in our training set. Although we areencouraged by our results to date, we still need to add moresamples to our training set and to improve the program's accuracy.<P>One of the key findings from our studies has been thediscovery that with our computer program, nuclear features (i.e.nuclear "grade") are more accurate in determining prognosis thanare determinations of prognosis based on the traditional measuresof tumor size and axillary node status. Our research may thus havea significant impact on the current practice of surgeon performingaxillary lymph node dissection at the time of mastectomy forprognostic purposes.<P>Presently, determination of whether breast cancer has spreadinto a patient's axillary lymph nodes is considered to be animportant factor in establishing prognosis: patients with tumorspread in axillary lymph nodes (e.g. node positive) have a higherlikelihood of tumor recurrence than those without tumor spread intoaxillary lymph nodes (e.g. node negative). Patients with nodepositive tumors tend also to have earlier recurrences and, ingeneral, do worse than patients with node negative tumors.<P>Data from our research, however, have indicated that apatient's nodal status (node positive vs. node negative) providesno new additional information about prognosis beyond that which isalready available from an FNA biopsy. When we use FNA data aloneto run our prognostic program, the results are just as accurate aswhen we run the program using FNA data plus information on nodalstatus and tumor size. That is, we have found that analysis of apatient's FNA sample alone provides all the information aboutprognosis that we need, and thus that performing an axillary lymphnode dissection solely for prognostic purposes is unnecessary.Should studies at other institutions confirm to our findings, manybreast cancer patients in the future could be spared the necessityof undergoing axillary lymph node biopsy.<H2>FUTURE DIRECTIONS</H2>We are now seeking to collaborate with other institutions tovalidate our computer program using their FNA samples. We have notpatented our program and are willing to provide copies of ourprogram to those researchers who sign collaborative agreements withus. We recently analyzed 19 breast FNA slides from UCLA and weregratified to find that we correctly diagnosed them all.<P>We are are also interested in expanding our computerizedinterpretative method to other organs besides breasts. We have hadgood early results with thyroid cancer and also believe ourapproach could prove particularly useful for evaluating lymphomas.<HR>Last modified: Thu Jan 4 16:45:23 1996 by Nick Street<ADDRESS> <!WA0><A HREF="http://www.cs.wisc.edu/~street/street.html">street@cs.wisc.edu</A></ADDRESS></BODY></HTML>⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?