Convolutional Neural Networks (CNN), a specialized form of the Artificial Neural Networks (ANN), is widely used in computer vision (CV) for a variety of object recognition applications such as the medical image classification. Depending on the number of classes and the characteristic properties of the classification problem, an appropriate activation function for each output layer of the CNN is constructed. When the softmax function is used in the output layer, it exponentially grows with its input and then saturates to its maximum value, which is usually assigned as 1. In binary classification, when the probability values are close to each other, classification success is expected to converge to zero since the CNN fails to recognize the patterns in the input-output training data. However, the softmax function produces a value that converges to 0.5. In this case, a transformation function must be applied to the output of the network. Thus, the classification success of the proportional output value is presented more realistically. In this paper, Mamdani Fuzzy Model is implemented to enrich the output of the binary classification problem of the computed tomography (CT) images with the CNN. Both CNN and fuzzy logic applications are implemented in Matlab environment. The results are extensively analyzed and compared to show the efficiency of the proposed approach in this paper. This method, which is proposed to interpret the output in binary classification problems, may be the subject of future studies as a supportive method in separating data belonging to overlapping classes in multi-class classification problems.
Anahtar Kelimeler: Binary classification, Convolutional neural network, Fuzzy logic, Image processing
|
