Abstract
Convolutional neural networks (CNN) are the best deep learning architecture to perform tumors classification for different imaging modalities: Us, X-ray, CT, and MRI. The scarcity of medical images and scarcity of resources are the contemporary problem for achieving successful classification. Therefore, it is preferable to use a simple network that does not require training and implementation resources than to use complex or pre-trained CNN models. Simple networks make it easy to use in clinical diagnosis and on mobile platforms. In this paper, a proposed CNN architecture for medical image multimodalities classification is presented. This proposed network is simple and is directly trained by medical images which is better than using pre-trained deep learning networks. Firstly, the data augmentation process is applied to avoid data shortage, and then, the proposed CNN is trained using the resulted augmented data. Simulation results demonstrate the efficiency of the proposed CNN architecture for efficient classification. The proposed model is trained on medical Us, X-ray, CT, and MRI datasets from scratch, and it can achieve 92.7%, 91.1%, 100%, 100% accuracies for these datasets, respectively.