This method thus allows the characterization of many tumors, because they are stiffer than normal tissues. Ultrasound elastography was previously used for the diagnosis of breast lesions [3], prostate cancer [4] and thyroid nodules [5]. However, the value of endoscopic ultrasound elastography for the diagnosis of pancreatic focal masses is not clear for the current moment, as some authors couldn’t differentiate benign and malignant pancreatic tumors [6]. Moreover, the intense fibrotic reaction and calcifications in chronic pancreatitis induce strain differences, and it is not clear if elastography is sensitive enough to detect them.

The study protocol is based on a semi-quantitative approach of EUS elastography data (movies) consisting of characterization of manually user-defined regions of interest, based on the hue histograms of the individual focal masses. Due to the inherent bias induced by selection of images from a dynamic sequence of EUS elastography, we have previously reported on the utility of using computer-aided diagnosis by averaging images from a dynamic sequence of EUS elastography [7]. A special plugin (based on the ImageJ software, NIH, Bethesda, MD, USA) was used to compute hue histograms on average EUS elastography images, while the hue histograms values for each patient (0 to 255 values) were further used to classify the patients with benign and malignant lesions.  

 

·        Patients diagnosed with solid pancreatic tumor masses, with cytological / histo­logical confirmation

·        Age 30 to 75 years old, men or women

·        Signed informed consent for EUS with elastography and FNA biopsy

·        Prior surgical treatment with curative intent or chemo-radiotherapy

·        Patients diagnosed with mucin producing tumors, pancreatic cystic tumors, etc.

·        Descriptive statistics

o       All results will be expressed as mean ± standard deviation (SD). Differences between the patients with pancreatic cancer and chronic pancreatitis will performed by the two-sample t-test (two independent samples). Since this parametric method makes assumptions about normality and similar variances, we will initially perform both the Kolmogorov-Smirnov and Shapiro-Wilk W normality tests and verify the equality of variances assumption with the F test. In the case of the two-sample t-test, we will also perform the non-parametric alternative given by the Mann-Whitney U test, since in some instances it may even offer greater power to reject the null hypothesis than the t-test.

o       Since with more than two groups of observations it is far better to use a single analysis that enables us to look at all the data in the same time, we will also perform the one-way analysis of variance (ANOVA) method with the same baseline assumptions. A p-value less than 0.05 will be considered as statistically significant.

·        Sensitivity, specificity, positive predictive value, negative predictive value and accuracy of EUS-EG will be determined in comparison with the final diagnosis

The use of the modern computer technology in medical decision support is now widespread and pervasive across a wide range of medical area. Thus, the Computer-Aided Medical Diagnosis (CAMD) is becoming an increasing important area for intelligent computer systems. Under these circumstances, there is a tremendous opportunity for the Artificial Intelligence, in general, and the Machine Learning, in particular, to assist the physicians deal with the flood of patient information and scientific knowledge. As a broad subfield of the Artificial Intelligence, the machine learningExploration and analysis by automatic means of large quantities of data in order to extract implicit, previously unknown and potentially useful information and discover meaningful patterns”). Machine learning can be viewed as an attempt to create a collaborative framework between human intuition and the huge computation power of the machines (computers). is concerned with the design and development of algorithms and techniques that allow computers to "learn". The major focus of the machine learning research is to extract information from data automatically, by computational and statistical methods and hence, machine learning is closely related to Data Mining (“

Common machine learning methods we shall use include:

·        Classification and decision trees;

·        Neural Networks;

·        Evolutionary algorithms

·        Support vector machines

 

            Classification and decision trees are used to predict membership of cases into two or more categories from their measurements on one or more predictor variables (making decisions algorithms). A distinctive characteristic of classification and decision trees is their flexibility. The ability of classification and decision trees to examine the effects of the predictor variables one at a time, rather than just all at once, is well-known. The ability of classification and decision trees to perform univariatesplits and examining the effects of predictors one at a time has implications for the variety of types of predictors that can be analyzed (categorical predictors, continuous predictors, or any mix of the two types of predictors).

            Classification and decision trees are not limited to univariate splits on the predictor variables. When continuous predictors are indeed measured on at least an interval scale, linear combination splits, similar to the splits for linear discriminant analysis, can be computed for classification trees. The classification and decision trees can be used in any classification of patients, decision concerning diseases, tissues, tumors types etc. The tree graph presents all the information in a simple, straightforward way, and probably allows one to digest the information in much less time than it takes for the traditional taxonomic techniques.

 

            The artificial neural network, commonly referred to as neural network (NN), is an information processing paradigm that is inspired by the way the human brain processes information. The key of this paradigm is the novel architecture of the information processing system, consisting of a large number of highly interconnected processing elements (neurons) working together to solve specific problems. This complex processing system, storing experimental knowledge and making it available for use, offers efficient mechanisms to extract patterns and detect trends that are too complex to be noticed by humans.

            There are two phases in neural information processing: the training phase and the testing phase. In the training phase, a training dataset is used to the progressive adjustments of the weighted interconnections (synaptic weights) that define the neural model. In this context, we speak of a training paradigm that refers to the process by which the free synaptic parameters of the NN are adapted to a process of stimulation by the environment in which the network is embedded. Afterwards, the trained neural model will be used in the testing phase to process testing patterns, yielding the true classification performance.

            Basically, NN is trained to associate outputs with input patterns. When NN is used, it identifies the input pattern and tries to output the associated output pattern. If a pattern that has no output associated with it is given as an input, NN gives the output that corresponds to a taught input pattern that is least different from the given pattern. NN will be used to learn how to classify patients, diseases, tissues, tumors etc., based on a training dataset of known cases.

 

            Evolutionary algorithms presume the existence of a population of candidate solutions (individuals) to a problem to be solved that must adapt to the task requirements (fitness assignment). The natural selection, the recombination and mutation mechanisms and the survival of the fittest triggers the average fitness of passing generations of individuals to rise step by step until the optimal solution(s) is (are) reached.

            In order to tackle diagnosis, evolutionary algorithms are applied under the form of an evolutionary classifier. The individuals are represented by decision rules that are evolved to fit the data samples and the aim is to achieve a complete rule set that models the decision making. The novel evolutionary classifiers are imagined in two distinct operating ways. The first direction involves the use of a multi-modal EA to discover the optimal rules for all classes of diagnosis. A second possibility regards the use of co-evolution to model the interaction between rules in solving the diagnosis task. On the one hand, population of rules, one for each diagnosis outcome, can cooperate towards completing an optimal rule set to fit to the data. On the other hand, a population of rules and a population of data samples can compete in an inverse fitness interaction that generates the evolution.

 

            Another way to target diagnosis through evolutionary algorithms is represented by hybridizations to classification-specific paradigms, such as support vector machines. The standard support vector machines regard diagnosis through the geometrical separation of positive and negative data samples. Despite the originality and performance of the learning vision, the inner training engine is intricate, constrained, rarely transparent and unable to offer convergence for any decision function. Therefore, an evolutionary resembling support vector machines technique is employed which considers the learning task as in support vector machines but uses an evolutionary algorithm to solve the optimization problem of determining the separation surface.

 

7.      Săftoiu A, Vilmann P, Ciurea T, Popescu GL, Iordache A, Hassan H, Gorunescu F, Iordache S. Dynamic analysis of endoscopic ultrasound (EUS) elastography used for the differentiation of benign and malignant lymph nodes. Gastrointest Endosc 2007; 66: 291-300.