Abstract
Background
Due to the lack of a ‘gold standard’ in the diagnosis of IBD the differential diagnosis between ulcerative colitis and Crohn′s disease can be very difficult. Verification of diagnosis of IBD takes a long time in majority of cases.
Methods
We have created an artificial neural network (ANN) of the multilayer perceptron type using the Neural Network Toolbox application from the MATLAB application package. Three types of images were used to train the ANN: the norm of the endoscopic picture of the colon, the endoscopic pictures of UC and CD. The first stage is the training of an artificial neural network to distinguish the presence or absence of pathology (29 images of the ‘normal colon’, 14 images of the CD, and 15 - UC). The second stage was to train the ANN to recognise the various forms of IBD. The network was trained on an array of 124 images (62 images of each class of pathologies). Each image was previously converted to the grayscale mode and then into a matrix of pixels. A vector with the number of elements equal to the size of the image was fed to the input of the perceptron.
Results
To solve the task of identifying the pathology a perceptron was built with 32,2784 input neurons, 10 hidden neurons and 2 output neurons which represent the conclusion that the image belongs to one of the two classes: norm or pathology. To solve the problem of differentiating CD and UC a perceptron was created with 364500 input neurons (this value was determined by the image resolution) and 2 output neurons representing the conclusion that the image belongs to one of the two classes: UC or CD. The best result in differentiation of pathology was shown by the ANN of MP 364500 type: 364500-20-2: 2, which total accuracy of recognition was 96,8%. The average accuracy of the developed model was 92.6%. However, in the control sample, the accuracy was 84.2%. This fact indicates that the model should be taught on more images. In addition to the ‘accuracy’ criterion, the ‘completeness’ parameter was used to evaluate the system. ‘Completeness’ for recognition of the image of the norm was the highest and equal to 1, for UC the criterion of ‘completeness’ was 0.89. The lowest ‘completeness’ was obtained when recognising the image of the CD (0.67).
Conclusion
ANN type MP 364500: 364500-20-2: 2 has shown the best results in the set targets. Efficiency in pathology recognition was 96.8%. The efficiency of the created ANN in solving the problem of recognition of different forms of IBD (UC/CD) can be described by the following parameters: specificity (Sp) −78.2%, sensitivity (Se) - 93.1%, accuracy (Ac) - 85,7%. The obtained ANN can be used to solve the problems of classification of endoscopic images of the intestine for the presence of IBD and for differential diagnosis.
Effect of an artificial neural network on radiologists' performance in the differential diagnosis of interstitial lung disease using chest radiographs.
