WO2001026026A2

WO2001026026A2 - Local diagnostic and remote learning neural networks for medical diagnosis

Info

Publication number: WO2001026026A2
Application number: PCT/US2000/027220
Authority: WO
Inventors: James K. Walker; Stephen A. Tuchman; Won Young Choi
Original assignee: University Of Florida
Priority date: 1999-10-04
Filing date: 2000-10-03
Publication date: 2001-04-12
Also published as: AU7749100A; WO2001026026A3

Abstract

The subject invention relates to a method and system for diagnosis of a medical condition. A specific embodiment of the subject invention utilizes a plurality of neural networks at a corresponding plurality of clinical sites to assist physicians in diagnosing a medical condition of a patient. Each of the plurality of neural networks can communicate with a server associated with a central neural network via, for example, the Internet. The server can receive patient data which can include, for example, images, patient information, parameters, biopsy information, and physician diagnoses. The central neural network can be trained on a large volume of medical cases, which come from the plurality of clinical sites. Even for relatively rare medical conditions, the neural network can be provided with diagnosed cases at an adequate rate for effective training of the central neural network. At appropriate times, the optimized parameters of the trained neural network can be transmitted to each of the plurality of neural networks at the clinical sites. The neural network at a site can thus assist a physician in reliably determining the nature and the likelihood of a medical condition even when it is dependent on a wide variety of patient data and even when the condition is relatively rare.

Description

DESCRIPTION

LOCAL DIAGNOSTIC AND REMOTE LEARNING NEURAL NETWORKS FOR MEDICAL DIAGNOSIS

Field of the Invention The present invention relates to medical diagnosis and more specifically to a system and method using neural networks for the diagnosis and interpretation of medical conditions.

Background of the Invention

The medical diagnosis task can be decomposed into three basic steps as follows: 1. Detection; 2. Classification; and 3. Recommendation. Detection refers to the step in which symptoms associated with one or more specific illnesses or conditions are first recognized. Classification is the process of designating or naming the condition, for instance, categorizing the condition into a known diagnostic group. Finally, recommendation is the step in which the physician prescribes a course of treatment for the condition.

The following problems are often encountered when performing one or more of these diagnostic steps m a typical clinical setting.

Consistency: On any given day, a physician may be fatigued or under stress. She or he may be inexpeπenced in a particular medical specialty. Identical clinical data and parameter values monitored for one patient may be interpreted differently by two physicians due to their different medical training, expeπence level, stress level, or other factors.

Transference/interpretation: One physician's mental rules in the diagnosis of a medical condition may be hard to descπbe and hence, difficult to transfer from one physician to another. These mental rules may also be difficult to explain to a patient if he asks how the physician arrived at the diagnosis or even to document reasoning for use by other physicians.

Nonlmeanty: When the relationships between the monitored values and the patient's condition are complex and not well understood, conventional, e.g., linear, statistical, models are often inaccurate and thus not sufficient or reliable. Therefore, diagnostic technology using more complex nonlinear models is clearly preferable and often necessary.

Infrequent Presentation: When a specific disease or condition is relatively rare, the physician has little expeπence in diagnosing the condition and may, therefore, misdiagnose.

These and other problems which are related at least m part to human errors and limitations m the area of medical diagnosis can be addressed successfully using computer-aided diagnostic tools. Conventional computer-aided medical diagnosis is based on statistical data analysis. More advanced diagnostic tools are based on artificial intelligence (Al) technology which generally involves expert systems, fuzzy logic, artificial neural networks, and various combinations thereof. The advent of effective, commercially available software and hardware tools of these types has greatly broadened the base of potential and realized medical applications. More recent examples of such use are disclosed in U.S. Patent Nos. 5,491,627,

5,486,999, 5,463,548, and 5,455,890. Still, none of the presently available medical diagnostic tools is capable of adequately addressing the problems discussed above.

Conventional computer-aided data processing techniques, such as linear regression, are difficult to implement successfully without well-defined relationships between the monitored values (inputs) and the patient condition (output). However, such well-defined relationships are seldom available especially because many medical conditions share common symptoms and are therefore difficult to detect and classify.

Expert systems represent a different Al approach in which complex systems are modeled using a set of Production Rules, i.e., IF/THEN rules. Expert systems are popular because of their design simplicity and their capability to recommend actions by inference or search. They have been shown to be beneficial m diagnosis problems under certain circumstances. However, the rule-based approach used in these systems requires a complete understanding of the task to be automated before an expert system can be implemented. Moreover, the large number of Production Rules required for increased robustness in the modeling of complex systems often slows down the decision-making process and aggravates maintenance due to the sheer number of rules to be kept track of

Fuzzy logic is typically used in situations where data and functional relationships cannot be expressed in clear mathematical terms. Instead, "fuzzy" relational equations are applied m which quantifiers such as "for many" or "for a few" are used to relate elements of different sets. Fuzzy logic systems provide conceptual advantages but require both intuition and expeπence m the proper design of working medical diagnosis systems.

Artificial neural networks ("neural networks") are networks of neuron-like units that can modify themselves by adapting to changing conditions. Unlike traditional Al systems which are rule-based, neural networks are very flexible and provide the capability of simulating complex nonlinear systems, the behavior of which is not well understood. This makes them uniquely suitable for medical diagnosis applications. Generally, neural networks mimic the ability of the human bram to recognize recurπng patterns on the basis of an inventory of previously learned patterns. In particular, they can predict the value of an output vaπable based on input from several other input vaπables that can impact it. The prediction is made by selecting from a set of known patterns the one that appears most relevant in a particular situation. Because of their flexibility m modeling complex systems, neural nets have been widely used in the medical practice.

Still, pπor neural networks have been limited in the extent of their training which requires an adequate amount of input data and corresponding reliable diagnoses. In particular, this is especially true and difficult to achieve for diagnoses of relatively rare diseases. The input parameters of such diseases may also be correlated with additional factors such as patient age, ethnic background, etc., which may be cπtical to the correct diagnosis of the patient's condition. Accordingly, such systems have provided limited ability for reliable diagnosis under these conditions. Thus, it can be seen that pπor art diagnostic tools based on classical statistical methods, expert system methods, and simple neural network methods have significant limitations when applied to medical diagnosis problems in general and especially to relatively rare medical diagnosis problems particularly where a disease or a medical condition is affected by a vaπety of patient-related additional parameters. Therefore, there is a need to develop a computer-aided medical diagnosis system and method that are capable of not only reliably determining the nature and the likelihood of a medical condition but also of providing an inteφretation dependent on a wide vanety of additional patient data even when the medical condition is relatively rare.

Bnef Summary of the Invention

Accordingly, it is an object of the present invention to provide a data processing system for medical diagnosis and inteφretation of medical conditions.

It is another object of the present invention to present a neural network system for estimating the likelihood of the medical condition on the basis of measurement data such as imaging data, interview data, and/or other input factors, and for interpreting the diagnostic output taking into account the contribution of vaπous input factors.

It is another object of the present invention to present a system for effectively training a neural network to diagnose a disease or medical condition even when the condition is relatively rare. It is yet another object of the present invention to present a system for effectively training a neural network to diagnose the pre-existing symptoms of a disease or medical condition even when the condition is relatively rare.

These and other objects are achieved m accordance with the present invention by providing a novel medical diagnosis system including a neural network for diagnosis, an Internet connection, and a server on which neural net training occurs. The neural network can be trained by being provided with the diagnosis made by a physician and with the measurement and interview data that was available to the physician. In case-by-case operation, the neural network system can use measurements and interview data to produce a score, or graded classification of the patient's medical condition to assist the physician in the diagnosis. In the present invention, it is envisioned that there exists a plurality of neural networks at a plurality of clinical sites. Each of the plurality of neural networks can then communicate with a central server via, for example, the Internet. The central server can then receive data from one or more of the plurality of clinical sites, including, for example, images, patient information, and physician diagnoses. By utilizing data from medical cases at multiple clinical sites, a neural network on the server can then be trained at a faster rate than an individual neutral network at one of the clinical sites could be trained utilizing only the data from medical cases at that clinical site. Accordingly, the server's neural network can learn at a fast rate, even for a relatively rare medical condition where there is not a sufficient volume of cases at an individual clinical site to effectively tram a neural network. Preferably, the central server receives data from many, if not all, of the clinical sites, such that there is an adequate rate for effective training at the central server. At appropπate times, the parameters of the central server's trained neural network are transmitted to each of the neural networks at the plurality of clinical sites.

In a preferred embodiment, the system can also provide an estimate of the likelihood of a medical condition being present which is characteπstic of the patient subsequently developing the disease. The neural network of the server can be trained for this task by receiving patient data, such as images and patient information, from the pπor records of patients who have been reliably diagnosed to subsequently have the disease.

Bnef Descπption of the Drawings

Figure 1 shows a schematic diagram of a system in accordance with the subject invention mcoφoratmg a plurality of neural networks at a plurality of clinical sites, each neural network connected via the Internet to a central server which houses a neural network.

Figure 2 shows a schematic diagram of a neural network having an input layer of processing elements, a middle layer of processing elements, and an output layer composed of a single processing element.

Figure 3 shows a block diagram of an embodiment of a data processing system for use m one of the plurality of clinical sites in accordance with the subject invention. Figure 4 shows a schematic diagram of an embodiment of a central server in accordance with the subject invention, mcoφoratmg dual servers where each contain framing neural networks with data storage facilities.

Figure 5 shows data and speculative curves for the performance of a neural network m detecting breast cancer m accordance with the subject invention.

Figure 6 shows a schematic diagram of a specific embodiment of the subject invention for the detection of breast cancer utilizing composite digital images.

Detailed Disclosure of the Invention The subject invention pertains to a method and apparatus for medical diagnosis. The subject invention can tram a centralized neural network to estimate the likelihood that a patient has a certain medical condition. This centralized neural network can be trained by data collected at a plurality of clinical sites, which communicate this data to a centralized server via, for example, the Internet. By utilizing data collected at a plurality of clinical sites, the centralized neural network is able to train much more quickly than if the training were based on data collected at an individual clinical site. Peπodically, the centralized neural network's parameters can then be transmitted to each clinical site to update the neural network at each site. The data acquired at each site can then be analyzed by the neural network on that site. The method and apparatus of the subject invention are particularly advantageous for medical diagnosis of relatively rare medical conditions, wherein it is difficult to have an adequate number of occurrences at an individual clinical site to effectively train a neural network.

The subject invention can be applied to medical imaging for the detection of certain medical conditions. Such medical imaging can include, for example, magnetic resonance imaging (MRI), CT scans, PET scans, ultrasound, x-ray, and other imaging techniques. In a specific embodiment, the subject method and apparatus can be utilized for the detection of breast cancer. For example, upon a reliable diagnosis of breast cancer in a patient, a radiographic image of a patient's breast tissue and other relevant data acquired at the individual clinical site can be communicated to a central server such that the image can be used to tram a centralized neural network. Other parameters can accompany the image, if desired. By receiving images from a plurality of sites, the centralized neural network can be trained more rapidly and effectively than a neural network at an individual site by images collected at that site The centralized neural network can then transmit updated parameters to each clinical site to update one or more neural networks at each site. This system can also be used for chest and other imaging for the detection of, for example, cancer. A. Description of Neural Net Diagnostic Hardware System. In the following descπption, like numbers designate like elements or processing steps as illustrated m the accompanying figures.

Figure 3 illustrates a processing system 10 for use at one of the plurality of clinical sites m accordance with the subject invention. Processing system 10 generally compπses a computer

12 which can be adapted to received input data from one or more sources, examples of which include an operator by means of a keyboard 14, a digital image file 24, and a patient database stored in memory 16. Memory 16 can also be used to store output data from the computer 12. Computer 12 can be coupled to display module 18 which may be a computer monitor or similar device. System 10 can further compπse a pπnter 19 for providing a hard copy of the diagnostic results. An interface 22, such as a modem, can be used for connection to a network of computers and/or the Internet.

Computer 12 can execute a simulation of a neural network 20 and an inteφreter unit 25. Computer 12 can be, for example, a PC, a mainframe computer, a server, or a workstation. Computer 12 can be connected via interface 22 to a Local Area Network (LAN), Wide

Area Network (WAN), or a packet switched network, such as the Internet. Accordingly, the information generated by processing system 10 can be made accessible to other computers via, for example, management software. In addition, processing system 10 can include vaπous other input/output (I/O) and peπpheral modules, as known m the art. Central Server. The subject invention can incoφorate a central server with a resident

"master" neural network. This master neural network can be trained by the input of appropπate data. A specific embodiment of the central server is shown schematically m Figure 4. The server shown in Figure 4 is connected to each of a plurality of neural networks located at a corresponding plurality of the clinical sites. The function of the neural networks located at the clinical sites is to assist the physicians in performing medical diagnoses. When a diagnosis is confirmed by biopsy or other reliable means, data corresponding to the patient can be sent to the central server. Such data can include, for example, relevant patient image data files, and relevant patient information.

The master neural network (NN) can be trained on the transmitted data sets to optimize the ability of the software to detect the presence of a given medical condition such as a particular disease. After a training peπod, the optimized parameters of the master NN can be downloaded to the NN at each of the plurality of clinical sites. The optimized, or updated, parameters received by the NN at each of the plurality of clinical sites updates the NN at each of the plurality of clinical sites to take advantage of the training of the master NN by data sets sent by one or more of the plurality of clinical sites. The master NN can also receive and be trained by data sets corresponding to a confirmed diagnosis which was not conducted at one of the plurality of clinical sites. Likewise, master NN can be enhanced or optimized by other framing if desired. In a specific embodiment, patient image data and information from the pπor year, or several pπor years, from an annual screening or checkup, from patients subsequently positively diagnosed, can be transmitted to the cenfral server. The data set can include the peπod of time between collection of the patient image and other data and the diagnosis of a certain stage of the medical condition. The cenfral NN can then be trained with respect to this data m order to optimize the ability of the software to detect the presence of symptoms characteπstic of the later development of a particular medical condition such as a disease. In other words, patient characteπstics which are correlated to, or predictive of, future development of a disease can be flagged, allowing preventive treatment and/or appropπate preparation. In a preferred embodiment, the master neural network algorithm can be optimized to detect the presence of a given medical condition and the detection of patient characteπstics which may indicate the likelihood of future development of the medical condition. The present invention is particularly advantageous with respect to a disease which occurs relatively rarely. By utilizing the confirmed diagnosed cases at a plurality of the clinical sites, an adequate number of cases can be obtained to effectively train the central NN. In this way, the present invention can achieve effective diagnostic neural networks at each clinical site even for relatively rare diseases. On the basis of the present disclosure, a person skilled in the art will also be able to use a central neural network for framing and improving the diagnostic functioning of neural networks at a plurality of sites.

Overview of neural networks. To fully appreciate the vaπous aspects and benefits produced by the present invention, a basic understanding of neural network technology can be useful. Following is a bnef discussion of neural network technology as applicable to the medical diagnosis system and method of the present invention.

Artificial neural networks loosely model the functioning of a biological neural network, such as the human bram. Accordingly, neural networks are typically implemented as computer simulations of a system of interconnected neurons. In particular, neural networks are hierarchical collections of interconnected processing elements configured, for example, as shown in Figure

2. Specifically, Figure 2 is a schematic diagram of a standard neural network having an input layer of processing elements, a middle layer of processing elements, and an output layer composed of a single processing element. The example shown m Figure 2 is merely an illustrative embodiment of a neural network 20 that can be used in accordance with the present invention. Other embodiments of a neural network can also be used Turning next to the structure of a neural network, each of its processing elements can receive multiple input signals, or data values, that are processed to compute a single output. The output value is calculated using a mathematical equation, known in the art as an activation function or a transfer function that specifies the relationship between input data values and the output value. Vanous parameters can be used to control, and therefore update, the relationship between input data values and the output value. As known in the art, the activation function may include a threshold, or a bias element. As shown in Figure 2, the outputs of elements at lower network levels are provided as inputs to elements at higher levels. The highest level element produces a final system output. In the context of the present invention, neural network 20 can be a computer simulation that produces a score, or graded classification, of a patient's medical condition, based on available measurements including, for example, image data file(s), responses, and other input factors. For instance, the scores produced by the network might range continuously from zero to one, with scores near zero indicating a low likelihood of disease and scores near one indicating a high likelihood of disease.

In the present invention, the neural networks in the plurality of clinical sites can be updated penodically such that they can be identical to the neural network in the central server. The neural network in the central server can continue to be trained, followed by additional updates of the clinical site neural networks. The neural networks in the plurality of clinical sites can then be used for providing scores for indicating likelihood of disease or other medical condition.

Training the Central Server. The central server can act as the learning neural network which is trained to diagnose a given medical condition.

With reference to Figure 4, the neural network can be framed by being provided with confirmed diagnoses made by physicians accompanied by input data such as image data, measurement data, and interview data, that was available to the physician. A given diagnosis along with the corresponding input data can be referred to as a data record. All available data records taken from a plurality of clinical sites compnse a data set. In accordance with the present invention, a data set corresponding to a particular medical condition can be stored m memory and made available for use by the processing system for training of the neural network m the central server.

A typical framing mechanism which can be used with respect to a preferred embodiment of the subject invention is bnefly descπbed. As known m the art, a myπad of techniques has been proposed in the past for framing feedforward neural networks. Most currently used techniques are vaπations of the well-known error backpropagation method. For further reference and more detail, the reader is directed to the excellent discussion provided by Rumelhardt, et al., m "Parallel Distπbuted Processing: Explorations m the Microstructure of Cognition," Vol. 1 and 2, Cambπdge: MIT Press (1986), and "Explorations in Parallel Distributed Processing: A Handbook of Models, Programs, and Exercise," both of which are mcoφorated herein by reference.

Bnefly, m its most common form, backpropagation learning is performed in three steps: forward pass; error backpropagation; and weight adjustment. As to the forward pass step, m accordance with a specific embodiment of the present invention, a single data record can be provided to the input layer of the network. This input data propagates forward along the links to the middle layer elements which compute the weighted sums and transfer functions as descnbed above. Likewise, the outputs from the middle layer elements can propagated along the links to the output layer element. The output layer element can then compute the weighted sum and transfer function equation to produce the patient score.

In the following step of the training process, the physician diagnosis associated with the data record can be made available. The score produced by the neural network can then be compared with the physician's diagnosis, which is expressed in mathematically comparable terms as a numeπcal score. Next, an error signal can be computed as the difference between the score corresponding to the physician's diagnosis and the neural network score. This error can then be propagated from the output element back to the processing elements at the middle layer through a series of mathematical equations, as known in the art. Thus, any error in the neural network score can be partially assigned to the processing elements that combined to produce it.

The outputs produced by the processing elements at the middle layer and the output layer can be mathematical functions of their connection weights. Errors in the outputs of these processing elements can anse from errors m the current values of the connection weights. Using the errors assigned at the previous step, weight adjustments can be made in the last step of the backpropagation learning method according to mathematical equations to reduce or eliminate the error in the neural network score.

The steps of the forward pass, error backpropagation, and weight adjustment can thus be performed repeatedly over the records in the data set. Through such repetition, the training of the neural network can be completed when the connection weights stabilize to certain values that minimize, at least locally, the diagnosis errors over the entire data set.

In addition to backpropagation framing, weight adjustments can be made in alternate embodiments of the present invention using different training mechanisms. It should be emphasized, that the present invention does not rely on a particular training mechanism. Rather, the only requirement is that the resulting network produce acceptable error rates m its sconng of patient conditions. Naturally, what is an acceptable error rate may in turn depend on the medical condition and other factors.

When the framing with the data set is completed, appropπate information can be transmitted to each of the plurality of neural networks. For example, in a specific embodiment, the values of the optimized connection weights can be downloaded from the server to each of the plurality of diagnostic neural networks at the plurality of clinical sites. The diagnostic neural network at each site can then proceed to perform its diagnostic function on new patients' input data. When a desired number of diagnosed cases, for example comparable to the pπor data set, has been reached at the plurality of clinical sites, the framing process can be repeated with respect to the central neural network. This framing can be done with the combined data sets to achieve a better set of optimized connection weights. Accordingly, this further optimized set of connection weights can then used by the diagnostic neural networks at each site for a period of time. The framing process can then be repeated with new data sets as they become available. Alternatively, the training can be terminated once a desired standard of accuracy is reached. In a preferred embodiment, the central server can have two servers, each with their own data storage as shown schematically in Figure 4. The neural network on server A can be trained on the data set resident in storage A, similarly for server B and storage B. The optimized neural network in server A can then be tested on the data set in storage B, and likewise for the other combination. The effectiveness of the framed neural networks can then be measured as known m the art by any one of a number of ways. An example is given in Example 1.

In a preferred embodiment of the present invention, when the database of diagnosed cases at the central server is large enough, "training expenments" can be performed to investigate correlation effects m the data. For example, two framing processes can be run for patient groups of two races. In this way, an optimized set of connection weights can be obtained for each race. Similar optimized connection weights can be obtained for patients in different age groups, lifetime estrogen exposure level, and a vanety of other parameters, some of which will be found to be important correlative factors for a given disease or medical condition. These optimized data sets which are now a function of certain patient-related parameters can be downloaded to the diagnostic neural networks. The neural net diagnosis can now be performed in an optimized manner tailored to the specific patient.

In another preferred embodiment of the present invention, a data set can be assembled for patient conditions corresponding to one or more specific time intervals pπor to the patient's diagnosis with a given disease. For example, data can be assembled corresponding to one year prior to a patients diagnosis with cancer. Neural network training can then be conducted on these "pπor year" data sets. Optimized connection weights can then be determined and downloaded to the diagnostic neural networks. Diagnosis can then be made for a patient condition symptomatic of disease developing in the subsequent year or other peπod of time. For such a diagnosis, the patient may be advised to return in six months for further tests. When enough data sets are assembled, the neural net framing can be performed for different patient- related parameters as before, thereby providing supeπor diagnostic quality.

It is often preferable to obtain very large numbers of patient data sets to achieve superior diagnostic objectives. Obtaining these large numbers can be achieved more quickly through the use of a training neural network m a central server which is connected to a plurality of diagnostic neural networks located m clinical settings. Output of Diagnostic Neural Network System. Upon completion of the neural network diagnostic process, the results can be displayed, for example on display 18, for use by the physician. The physician can review the results to aid m his or her diagnosis of the patient condition. The displayed results can also be pπnted, for example on pπnter 19, to create a record of the results of the neural network process. In the case where part of the patient input data was in the form of an ιmage(s), the pπnted output can be in the form of a reproduced image with supeπmposed indicator marks for the physician's consideration which are indicative of possible specific disease or different diseases at each of these locations.

In addition, interface 22 can permit communication of the results to other physicians. Most importantly, interface 22 can permit Internet, or other equivalent, communication with the central server in the manner descnbed in the present application.

Extensions and Applications. The diagnostic system and method of the present invention were descnbed with reference to a specific application which is the use of the invention for the diagnosis of medical conditions. It should be clear, however, that the pnnciples of this invention that provide for an enhanced inteφretive facility due to increased training even for highly specific and even relatively rare circumstances. For example, the present invention can readily be applied in areas as diverse as financial analysis, electronics design, oil exploration, fatigue determinations, and others. In particular, the inteφretation of diagnostic scores provided by the present invention can be used in vanous complex systems for the puφoses of prediction, planning, monitonng, debugging, repair, and instruction. More specifically, results from the inteφretation of systems scores obtained using this type of super- framed neural network can be used to develop production rules as part of an expert system, or to provide further insight into fuzzy relationships used m other artificial intelligence systems. The following examples illustrates the use of the system and method of the present invention for the diagnosis of a particular medical condition. Example 1 — Application to Breast Lesion Diagnosis

Mammographic annual screening has proven to be an effective method for early detection of breast cancer. Independent double reading of the images by two radiologists has proven to increase the sensitivity of screening. Computer-aided diagnosis (CAD) has recently been shown to be similarly effective as a second radiologist increasing the sensitivity of screening.

Many researchers have developed algorithms for computenzed analysis of mammograms. Methods, including the framing of neural networks, have been developed to detect spiculated lesions, masses, micro-calcifications, and other pathologies. It is not necessary here to review the details or relative advantages of the different neural network algonthms developed for these puφoses.

Following is a bnef review of how the above neural networks have been typically used in practice. A CAD workstation hosting the neural network is located at a given mammographic facility. The digitized film or digitally acquired image is processed by the neural network and a CAD output is made available for consideration by the radiologist. The connection weights used by the neural network have been determined by the manufacturer of the CAD system based on a sample of one hundred to at most a few hundred cases of breast cancer. Considering that there are several different pathologies such as spiculated masses, micro-calcifications, masses, etc., there is a very limited number of cases of specific pathologies for training puφoses. This results in neural network training which is less than optimal and less than optimal sensitivity for breast cancer detection. In addition, due to the limited number of cases for NN framing, there is little or no opportunity to correlate patient parameters such as age, race, lifetime estrogen exposure level, etc., which are known to correlate with the presence of breast cancer.

The reason for the very limited number of cases available for NN framing is that the NN algorithm development and framing has always occurred at a single clinical site, frequently a university hospital. Such a site may screen about 10,000 patients per year among whom there are less than 100 cases of breast cancer. Thus, over the course of a few years of research and development, there are never more than a few hundred diagnosed cases available for NN framing. The National Institute for Cancer has attempted without success to create national digital databases of breast cancer cases.

Figure 5 shows data for the performance of a typical neural network in detecting breast masses. The true positive (TP) detection rate is plotted versus the number of diagnosed cases used for framing the neural network. There are only two data points, at slightly over 90% TP and slightly over 80% TP, with a database of 253 diagnosed cancer cases. The two data points correspond to a false positive (FP) rate per image of 4.2 and 2.0 respectively. Needless to say, it is desirable to keep the FP rate as low as possible. Two curves with arbitrary but reasonable shape are drawn through the two data points. These curves indicate the general anticipated performance of the neural network versus the number of diagnosed cases used for training the neural network. There is no theoretical or expenmental information on the detailed shape of these curves. Nevertheless, it is a fact that for each 1% increase in True Positive detection sensitivity by a better trained neural network, 2,000 additional patients can be diagnosed with breast cancer. This illustrates the importance of making available a major increase in the number of diagnosed cases for framing the neural network which is a major objective of this specific example of the present invention The present invention can utilize a central server whose resident NN is trained by diagnosed cases from a multiplicity of clinical sites. Due to the large and ever-increasing number of diagnosed cases, the NN diagnosis can also be trained for subsets of patients grouped according to disease-related parameters as discussed earlier. In addition, NN training can be performed for pnor year mammographic images. Pathology of breast disease prior to cancer detection can then be identified as descnbed earlier. According to the present invention, major and novel benefits can result from having a very large number of disease-diagnosed patient data available for neural network training. Furthermore, according to the present invention, a means is disclosed to effectively and efficiently achieve this objective.

Example 2 — Application to Breast Lesion Diagnosis Using a Dual Film Cassette

The traditional annual mammographic screening m Example 1 is performed with a cassette containing a single screen and single film. Screening can also be performed with a cassette containing a single screen and two films located on the opposite faces of the screen as disclosed in U.S. Pat. Nos. 5,751,787, PCT/US97/15589 and pending U.S. Serial No. 09/075,670, both of which are mcoφorated herein by reference. The first film is exposed m the usual way as in Example 1. The second film may be exposed to a fraction of the light intensity compared to the light intensity exposing the first film. Preferably, the second film is essentially identical to the first film, and is exposed to about half the light intensity compared to the light intensity exposing the first film. In this way, the effective latitude of the dual film system is about twice that of the conventional single film technique. This increased latitude overcomes the major limitation of conventional film screen mammography. The two films may be digitized m a commercial film digitizer and the two digital image files manipulated and combined, if desired, in a computer to produce a single digital image file with wide latitude as disclosed in U.S. Pat. Nos. 5,751,787, PCT/US97/15589 and pending U.S. Seπal No. 09/075,670. Figure 6 shows a schematic of a specific embodiment of the subject invention where two images are used to form a composite digital image which is then analyzed by one of a plurality of neural networks. Each of the plurality of neural networks received periodic updates from a central neural network.

The neural network system may be applied to analyze each digital image separately and/or the combined image to detect pathologies as descnbed m Example 1. All other considerations discussed m Example 1 are applicable in this example.

Example 3 — Lung Lesion Diagnosis Using a Digital X-Ray Image Acquisition System

The method and apparatus of the subject invention can also be utilized for in x-ray imaging of the chest to diagnose the presence of lung lesions or other abnormalities , the process of diagnosis by the radiologist can be difficult and time consuming, even with an advanced digital image acquisition system. The signal of the presence of a lesion may be very subtle and easily "missed" especially when "hidden" behind the portions of the image occupied by bones In these cases, the subject NN system can be efficiently framed to identify these lesions. The subtlety of these diagnoses is such that great benefit can be obtained from the large number of cases utilized to tram the central neural network.

Although the foregoing descnption and Examples refer to particular preferred embodiments, it will be understood that the present invention is not so limited. It will occur to those of ordinary skill in the art that vaπous modifications can be made to the disclosed embodiments, and such modifications are intended to be withm the scope of the present invention which is defined in the following claims.

Claims

1. A method for diagnosis of a medical condition, comprising the following steps: (a) providing input data with respect to a patient; (b) converting said input data into numerical data; (c) inputting said numencal data to a first neural network to diagnose a medical condition on the basis of the input numeπcal data, wherein said first neural network receives parameters from a second neural network, wherein said second neural network is trained with data from patients who have been diagnosed to have the medical condition, and wherein inputting said numencal data to said first neural network produces at least one score relating to the diagnosis of the medical condition.

2. The method according to claim 1, wherein the diagnosis of the medical condition is the likelihood of the patient having the medical condition.

3. The method according to claim 1 , wherein said second neural network is trained with data from patients acquired approximately a first time penod pπor to said patients being diagnosed to have the medical condition, and wherein the diagnosis of the medical condition is the likelihood of the patient developing the medical condition within approximately a second time penod.

4. The method according to claim 1 , wherein said first neural network is one of a plurality of neural networks located at a corresponding plurality of locations, each receiving said parameters from the second neural network and each able to receive said input numencal data and produce at least one score relating to the diagnosis of the medical condition.

5. The method according to claim 4, wherein said second neural network is associated with a server which is connected to the plurality of neural networks via a connectivity means selected from the group consisting of: the Internet, a wide area network, and a local area network.

6. The method according to claim 4, wherein said plurality of neural networks are implemented by a corresponding plurality of processors, wherein said plurality of processors can each send the server associated with the second neural network data corresponding to patients who have been diagnosed to have the medical condition.

7. The method according to claim 6, wherein the second neural network is framed on data received from the plurality of processors.

8. The method according to claim 1, wherein the second neural network is framed on a set of data corresponding to patients having a given characteristic.

9. The method according to claim 8, wherein the charactenstic is age interval.

10. The method according to claim 8, wherein the characteristic is race.

11. The method according to claim 8, wherein the characteristic is high genetic πsk.

12. The method according to claim 8, wherein the characteristic is hormonal history.

13. The method according to claim 8, wherein the charactenstic is family history of the medical condition.

14. The method according to claim 8, wherein the charactenstic is reproductive history.

15. The method according to claim 6, wherein one of said plurality of processors sends the server associated with the second neural network data corresponding to at least one time that the patient was seen by a physician pnor to the time that the patient was diagnosed to have the medical condition.

16. The method according to claim 1, wherein the medical condition is breast cancer.

17. The method according to claim 1 , wherein the input data compπses a digital image file of a mammogram.

18. The method according to claim 17, wherein the digital image files have been obtained from a single film technique used in screen film mammography.

19. The method according to claim 17, wherein the digital image file of a mammogram has been obtained by a direct digital acquisition system.

20. The method according to claim 17, wherein the input data further compπses a second digital image file of a mammogram wherein said digital image files correspond to mammograms based on simultaneously exposed films.

21. The method according to claim 20, wherein the two digital image files may be processed by the neural network separately.

22. The method according to claim 20, wherein the two digital image files may be combined into a composite digital image file before being processed by the neural network.

23. The method according to claim 20, wherein said simultaneously exposed films are exposed by a single screen.

24. The method according to claim 1, wherein the input data compπses a digital image file of a chest x-ray.

25. A system for diagnosis of a medical condition, comprising: a first neural network for receiving numencal data, wherein said numencal data is generated by converting input data with respect to a patient into said numencal data; and a second neural network, wherein said second neural network is trained with data from patients who have been diagnosed to have the medical condition, wherein said first neural network receives parameters from said second neural network, and wherein the input of said numeπcal data to the first neural network produces at least one score relating to the diagnosis of the medical condition.

26. The system according to claim 25, wherein said first neural network is one of a plurality of neural networks located at a corresponding plurality of locations, each receiving said parameters from the second neural network and each able to receive said input numencal data and produce at least one score relating to the diagnosis of the medical condition.