Future Generation Computer Systems 15 (1999) 287–292
Computer implementation of a medical diagnosis problem by pattern classification Domenico Conforti ∗ , Luigi De Luca 1 Department of Electronics Informatics and Systems, University of Calabria, 87030 Rende CS, Italy
Abstract In this paper we present a software system which can aid the medical diagnostician for the diagnosis of breast cancers. The system has been developed on a “Windows 95” platform and provides a user friendly interface, made up of windows and visualization tools. An interesting and innovative feature is represented by the telemedicine configuration of the software system, which can be run in a remote fashion, exploiting, from some remote regions, the expertize and the clinical database available in advanced medical centers. A prototype version of the software system, named CAMD (computer aided medical diagnosis) is currently being tested and validated with the collaboration of the Cytopathology Department of the Cosenza General Hospital (Calabria, Italy). c 1999 Elsevier Science B.V. All rights reserved.
Keywords: Medical diagnosis; Pattern classification; Mathematical programming; Telemedicine; Web application
1. Introduction The main aim of this paper is to present a prototype software system which could aid the physician during the diagnosis process of some serious diseases, on the basis of a small amount of clinical information. Important issues motivate this work: the availability of methods and technologies, which can be implemented as computerized automatic system, could increase the objectivity and the speed of the diagnosis process; a computerized diagnostical system could assist the physician in making a more reliable diagnosis, even in case the clinical tests are inaccurate; a computerized diagnostical system could easily and effectively be embedded into a telemedicine system for advanced diagnostical teleconsults (remote diagnosis). Moreover, it is worthwhile to mention that the availability of several advanced technologies (e.g. image processing, computer vision, inductive learning, mathematical programming, neural networks, parallel computers, telecommunications) represents a driving force for improving and making more efficient the medical diagnosis of several serious diseases. In fact, we have combined some of these technologies (image processing, inductive learning, mathematical programming, telecommunications) with the aim to develop a software tool which can aid and support the physician ∗ 1
Corresponding author. E-mail:
[email protected] E-mail:
[email protected]
c 0167-739X/99/$ – see front matter 1999 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 7 3 9 X ( 9 8 ) 0 0 0 7 3 - 9
288
D. Conforti, L.D. Luca / Future Generation Computer Systems 15 (1999) 287–292
in several activities: diagnosis of diseases, diagnostic consultations, patient scheduling, physician-to-hospitals and hospitals-to-hospitals links. In principle, the medical diagnosis process can be interpreted as a decision making process, during which the physician induces the diagnosis of a new and unknown case from his/her specific clinical experience. The physician’s clinical experience can be viewed as the knowledge base with an appropriate learning rule on the basis of which he/she makes the diagnosis. The medical diagnosis problem considered in this paper is the diagnosis of breast cancers, that is, to discriminate benign or malignant lumps in the breast. In this case, the diagnosis problem can be modeled as a pattern classification problem, since the process is to decide a class membership, the class of benign cases versus the class of malignant cases. Pattern classification is a fundamental problem which lies within the field of inductive learning, a machine learning methodology on the basis of which it is possible to learn general knowledge from some specific cases. In this respect, artificial neural networks (ANNs) represent well-known tools, typically effective in the implementation and solution of pattern classification problems. In fact, ANN has been used for several medical diagnosis problems, in the case of different diseases. A different approach for implementing and solving pattern classification problem is based on the use of mathematical programming methods [2]. In fact, several authors showed that to discriminate among general and different patterns is equivalent to finding nonnegative solutions of linear equalities [5], or to find optimal solution of linear and nonlinear optimization problems [1,7,8]. We followed one of these new approaches [10] for developing our prototype software system. The paper is organized as follows. In Section 2, we present the basic details of the specific medical diagnosis problem considered; in Section 3, we illustrate the computer implementation and the development of the software system; some concluding remarks and further developments complete the paper in Section 4.
2. The medical diagnosis as pattern classification problem: An example As an example of medical diagnosis problem, we consider the early detection of breast cancers, on the basis of a fine needle aspirate (FNA) sample, needle biopsies which provides a way to examine a small amount of tissues from a breast lump [3]. Using a microscope, by visual interpretation of some morphological characteristics of cell nuclei, well-experienced physicians are able to successfully diagnose breast cancers on the basis of FNA. On the other hand, the ability of FNA procedures to correctly diagnose cancer varies widely; generally, the diagnosis process is highly subjective, depending upon the visual and empirical interpretation of the physician [4]. However, it is possible to increase the objectivity and reliability of FNA diagnosis by coupling image processing techniques with machine learning methods, more specifically following a pattern classification approach. In fact, it is worthwhile to observe that the diagnosis process based on FNA is a potential candidate to be computerized, in order to make more objective and reliable the evaluation of the clinical data. In order to develop a computerized diagnostical system, we have to replicate, to some extent, the medical expertize, that is, to istantiate the problem solving ability of the medical diagnostician into a software system.
3. Computer implementation The computerized diagnostical system, proposed in this paper, is basically composed of an image processing software package for the interactive processing of the digitized image from the microscope slide of FNA, and by an automatic classifier based on inductive learning, which uses a linear programming model to discriminate among elements belonging to different sets (see Fig. 1).
D. Conforti, L.D. Luca / Future Generation Computer Systems 15 (1999) 287–292
289
Fig. 1. CAMD architecture
3.1. Image features extractor The image processing software package implements a graphical computer program for the measurement of cytological features, based on a digital scan of FNA samples. An FNA sample is prepared by taking cellular fluid from the patient’s breast, placed on a glass slide, stained to highlight the nuclei of the cells and mounted on a microscope. A portion of the slide, in which the cells are well-differentiated, is then scanned using a video camera and a frame-grabber board. The digital image is then stored as a file, in a suitable format, onto a PC and it is made available on the PC monitor in order to be processed. In order to measure some parameters related to size, shape and texture of the nuclei, an image segmentation program is used to determine the exact boundaries of the nuclei; by a mouse pointer, the user draws the approximate boundary of each nucleus, and by an image processing approach known as “snake” [6], these approximations then converge to the exact nuclear boundaries. Once all or most of the nuclei have been isolated in this fashion, the program computes 10 features for each nucleus: area, radius, perimeter, symmetry, number and size of concavities, fractal dimension of the boundary, compactness, smoothness and texture. Then, the mean value, extreme value and standard error of each of these cellular features are computed for each image, resulting in a total of 30 real-valued nuclear features for each FNA sample. Therefore, it is possible to generate a 30-dimensional features vector for each patient. Performing this analysis for each individual on a large set of patients, for which the actual diagnostic outcome is known, it is possible to generate a training set (a set of 30-dimensional feature vectors) with which a classifier can be constructed to diagnose future and new FNA samples, that is, to discriminate between benign and malignant samples. 3.2. Classifier The automatic classifier is based on a linear programming model in order to discriminate, in the R30 space, among points related to benign cells and points related to malignant cells. Following the same notation of Mangasarian et al. [9], we sketch the mathematical structure of this classifier. Given n = 30 (the total number of nuclear features), let p malignant n-dimensional vectors be stored in the p ×n matrix A, and q benign n-dimensional vectors be stored in the q × n matrix B. With the aim to discriminate points in A and B, we can construct a plane in the n-dimensional real space Rn , which can be mathematically represented by wT x = γ , where x is the generic element in Rn and T is the vector transpose operation. It is easy to see that this plane is able to strictly separate points in A from points in B, if and only if the following inequalities are satisfied: Aw ≥ γ e + e,
Bw ≤ γ e − e,
where e is a vector of ones of appropriate dimension. In general, the two sets A and B are not strictly separable by just one plane and the above inequalities cannot be satisfied. Therefore, we can only satisfy the inequalities in some approximate way, for instance, by minimizing the average sum of their violations.
290
D. Conforti, L.D. Luca / Future Generation Computer Systems 15 (1999) 287–292
In fact, let y and z be two real vectors, with dimensions p and q, respectively, for which each component represents the violation of each inequality. Then, the above inequalities can be rewritten in the following way: Aw + y ≥ γ e + e,
Bw − z ≤ γ e − e.
We can find the parameter w and γ (defining the separating plane and satisfying the above inequalities), by trying to reduce the contribution of y and z. This can be done by minimizing the average sum of all violations, that is, by solving the following linear programming problem: T eT z e y + , min w,y,z,γ p q subject to Aw + y ≥ γ e + e, Bw − z ≤ γ e − e, y, z ≥ 0. The solution of this problem gives a strict separating plane that satisfies the inequalities, if such a plane exists (in this case y = z = 0); otherwise, it minimizes the average sum of the violations y and z of the inequalities. In case there is presence of an unacceptable mixture of benign and malignant points in the halfspaces generated by the separating plane, the same procedure can be applied recursively to one or both of the halfspaces. Once all possible separating planes are generated, a class membership in Rn has been determined, defined by the regions in Rn in which there are points of A and points of B. A new and unknown case can be easily discriminated by evaluating its position with respect to the separating planes. In order to make the above approach useful in a clinical environment, the physician needs some idea of “how benign” or “how malignant” a new case is. By the data coming from the classification phase, it is possible to construct a statistical method for computing the approximate probability of malignancy. In this way, the physician has useful information on the basis of which to decide a therapeutic treatment; furthermore, the patient can easily compare the diagnosis with respect to the other cases belonging to the training set, in much the same way that an experienced physician takes advantage of years of experience. 3.3. Telemedicine and world wide web configuration With the aim to provide a telemedicine configuration of the computerized diagnostical system, we have developed a different version of CAMD based on the client/server paradigm, in order to allow remote diagnosis (see Fig. 2). In fact, the server is able to accept queries from the client, execute the automatic classification stage and make the diagnosis, then send the results to the client. On the other hand, the client is able to extract and measure the cytological characteristics from the FNA sample, send these measures to the server, receive the diagnosis from the server and locally visualize the results of the diagnosis. By this way, it is possible to collect in an advanced medical center (where the server is installed) a large amount of tested clinical data, on the basis of which developing a suitable training set, and consequently, a robust automatic classifier, guarantees a more reliable and accurate diagnosis. These potentialities are made available for all the clients connected to the server, including those located in a remote and isolated region, where, typically, an advanced expertize and a suitable clinical database are not available in the specific field of cancer diagnosis. A further development of CAMD has been its integration into the world wide web environment. We are currently testing two different technical approaches that are based on CGI (common gateway interface) applications and Java web-based applications. In any case the aim is to implement the following functionality: to allow remote diagnosis from any site (physician’s house, remote medical center, etc.) where an internet connection and a web browser exist.
D. Conforti, L.D. Luca / Future Generation Computer Systems 15 (1999) 287–292
291
Fig. 2. Telemedicine configuration
From the user point of view, both approaches are equivalent. In fact, the physician accesses a protected medical image archive, hosted by a secure web server of a central hospital, and selects, by a web browser, the relevant image. Then he selects some suspected cells and activates the diagnosis process by mouse, as in the stand alone version of CAMD. On the other hand, the two approaches differ in terms of computational workload. In the first case, image processing and diagnosis computation are performed by the remote server, which can turn out to be overloaded. The Java approach, on the contrary, is based on local execution of Java applets. In practice, the remote web server sends to the physician’s web browser the selected patient’s image and an applet to process it; the whole computation, and in particular the image processing and classification procedure, is demanded to the web browser, while the server is free to manage other requests from physicians in different sites.
4. Conclusions The computerized diagnostic system, sketched in this paper, is currently at the level of prototype. Further developments, improvements and extensions are under investigation and are based on the following issues: – improve the performance of the computerized diagnostic system with the aim to reduce the bias during the selection of the cells to be digitized and to make more precise the identification of the cell boundaries; – improve the performance of the classifier, in order to make more robust the classification phase, by using, for instance, a different approach based on nonlinear optimization models; – extend the application of the automatic diagnostic system to other types of cancers (for instance brain and thyroid cancers) and other types of diseases; – tuning the automatic diagnostic system on a large set of real cases, interacting with end-user medical institutions.
Acknowledgements We would like to thank the cytopathology department of Cosenza General Hospital for the contribution in doing this research activity. This research was financially supported by European Commission DGIII/B under
292
D. Conforti, L.D. Luca / Future Generation Computer Systems 15 (1999) 287–292
the ‘Euromed’ project and by the Italian Ministry of University within the special project ‘MOST: Methods of Optimization for Systems and Technologies’, 1997–98. References [1] K.P. Bennett, O.L. Mangasarian, Robust linear programming discrimination of two linearly inseparable sets, Optimization Methods Software 1 (1992) 23–34. [2] G.B. Dantzig, Linear Programming and Extensions, Princeton University Press, Princeton, NJ, 1963. [3] W.J. Frable, Thin-needle aspiration biopsy, in: Major Problems in Pathology, Saunders, Philadelphia, PA, 1983. [4] R.W.M. Giard, J. Hermans, The value of aspiration cytologic examination of the breast, A statistical review of the medical literature, Cancer 69 (1992) 2104–2110. [5] W. Highleyman, A note on linear separability, IRE Trans. Electron. Comput. (1961) 777–778. [6] M. Kass, A. Witkin, D. Terzopoulos, Snakes: active contour models, Int. J. Comput. Vision 1 (4) (1988) 321–331. [7] O.L. Mangasarian, Linear and nonlinear separation of patterns by linear programming, Oper. Res. 13 (1965) 444–452. [8] O.L. Mangasarian, Multisurface method of pattern separation, IEEE Trans. Inform. Theory 6 (1968) 801–807. [9] O.L. Mangasarian, W.N. Street, W.H. Wolberg, Breast cancer diagnosis and prognosis via linear programming, Oper. Res. 43 (1995) 570–577. [10] O.L. Mangasarian, W.H. Wolberg, Cancer diagnosis via linear programming, SIAM News 23 (1990) 1–18.
Domenico Conforti is Assistant Professor in Operations Research at the Department of Electronics, Informatics and System, University of Calabria, Italy. He received “Laurea” degree in Electrical and Computer Engineering from University of Calabria in 1985. In 1985 and 1986, he was research fellow at the University of Karlsruhe, Germany, and at the Mathematics Research Center, University of Wisconsin, USA. In 1988 he was appointed as researcher at the European Center for Scientific and Engineering Computing (ECSEC) in Rome. In 1993, he received an honorary fellow at the Center for Mathematical Sciences, University of Wisconsin, USA, where he worked in the field of parallel algorithms for the solution of large scale decision making problems. He is co-author of more than 50 papers on numerical algorithms for nonlinear optimization, computational engineering, parallel algorithms and software for vector and parallel computing. His current research interests include parallel algorithms for nonlinear optimization problems and application of optimization models and methods to the solution of classification problems. Luigi De Luca is a research associate at Parallel Computing Laboratory of University of Calabria (Italy); system manager of the minisupercomputer Alliant FX/80 (1990–93); member (1990–94) of the Research Team of the Project “Sistemi Informatici e Calcolo Parallelo”, National Research Council of Italy; member (1992) of the Research Team of the Project “Transporti 2”, National Research Council of Italy; member of the Research Team of the Project EUROMED (1995–98), European Commission; member (1997) of the Research Team of the Project EUROMED ETS, European Commission. He currently holds an acting associate professorship in the field of simulation techniques, at the Department of Electronics, Informatics and Systems of University of Calabria. His research interests include: parallel computing systems, neural networks, numerical simulation techniques, telemedical systems. He is author of several papers in the field, published by well-established journals.
