📄 http:^^www.cs.wisc.edu^~street^abstract.html
字号:
Date: Tue, 05 Nov 1996 00:25:59 GMTServer: NCSA/1.5Content-type: text/htmlLast-modified: Mon, 05 Sep 1994 21:28:42 GMTContent-length: 2639<html><head><title> Abstract of Nick Street's Thesis </title></head><body><h1> CANCER DIAGNOSIS AND PROGNOSIS VIA LINEAR-PROGRAMMING-BASED MACHINE LEARNING <p>W. Nick Street<br>Under the supervision of Professor Olvi L. Mangasarianat the University of Wisconsin-Madison</h1><br>The purpose of this research is twofold. The first purpose is thedevelopment of machine learning methods based on linear programming. These advances come in the form of both novel learning algorithms andgeneralization-improvement approaches that are applicable to a widerange of learning algorithms.The second aim is the application of these algorithms, along withideas from the fields of statistics and image processing, to problems arising inthe diagnosis and treatment of breast cancer. The first chapter provides background into the relevant areas ofresearch, most importantly machine learning andbreast cancer diagnosis and prognosis, and describes previous workthat has united the two fields.<p>The first major contribution of this research is anautomated cytological analysis system that we call <b>Xcyt</b>. Thissystem includes a partially automatic image segmentationfacility for isolating cell nuclei from a digital image.<b>Xcyt</b> allows computation of various size, shape and texturefeatures of these nuclei. A linear separator distinguishes benigncases from malignant cases based on 569 training examples. <b>Xcyt</b>also provides an estimate of the probability of malignancy for eachcase. The <b>Xcyt</b> system is currently in use at the University ofWisconsin Hospitals.<p>We also address the problem of cancer prognosis, that is, determiningwhen a cancer is likely to recur. A new learning method, therecurrence surface approximation (RSA), is introduced for predictingthe time of recurrence. This procedure uses linear programming topredict recurrence from right-censored input data. Various extensionsand variations of RSA are also explored, including a separation-basedprocedure that we call implicit RSA (IRSA).<p>The topic of improving the generalization of a learning system isfirst addressed in the context of RSA, where we introduce a procedure for choosing the most relevant input features foruse the the predictive model. The final two chapters also describegeneralization-enhancement methods. The first is mock generalization,which incorporates a tuning set into the optimization process. Thesecond is banded approximation, which explicitly defines a toleranceband around the predictive surface in order to avoid overfitting thetraining data.</body></html>⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -