A New Convolution Neural Layer Based on Weights Constraints

With the rapid development of artificial intelligence techniques, emerging deep neural networks (DNN) is one of the most effective ways to solve many challenges. Convolution neural networks (CNNs) are considered one of the most popular AI techniques used to extract and analyze meaningful features for image datasets, especially in the medical diagnosis field. In this paper, a proposed constrained convolution layer (COCL) for the CNN model is proposed. The new layer uses a constrained number of weights in each kernel trained in the phase of learning and excludes the others weights with zero values. The proposed method is introduced to extract a special type of feature considering the local shape of a sub-image (window) and topological relations between group pixels. The features extract according to a random distribution of weights in kernels that are determined considering a particular desired percentage. Furthermore, this paper proposed a CNN model architecture that uses COCL rather than the traditional CNN layer (TCL). The efficiency of the method is evaluated using three types of medical image datasets compared with the traditional convolution layer, pre-trained deep neural networks (pre-DNNs), and state-of-art methods. The proposed model outperforms other methods in terms of accuracy and F1 score metrics and exceeds more than 98%, 89%, and 93% for the three datasets used in the evaluation, respectively.