Abstract
Effective medical diagnosis is dramatically expensive, especially in third-world countries. One of the common diseases is pneumonia, and because of the remarkable similarity between its types and the limited number of medical images for recent diseases related to pneumonia, the medical diagnosis of these diseases is a significant challenge. Hence, transfer learning represents a promising solution in transferring knowledge from generic tasks to specific tasks. Unfortunately, experimentation and utilization of different models of transfer learning do not achieve satisfactory results. In this study, we suggest the implementation of an automatic detection model, namely CADTra, to efficiently diagnose pneumonia-related diseases. This model is based on classification, denoising autoencoder, and transfer learning. Firstly, pre-processing is employed to prepare the medical images. It depends on an autoencoder denoising (AD) algorithm with a modified loss function depending on a Gaussian distribution for decoder output to maximize the chances for recovering inputs and clearly demonstrate their features, in order to improve the diagnosis process. Then, classification is performed using a transfer learning model and a four-layer convolution neural network (FCNN) to detect pneumonia. The proposed model supports binary classification of chest computed tomography (CT) images and multi-class classification of chest X-ray images. Finally, a comparative study is introduced for the classification performance with and without the denoising process. The proposed model achieves precisions of 98% and 99% for binary classification and multi-class classification, respectively, with the different ratios for training and testing. To demonstrate the efficiency and superiority of the proposed CADTra model, it is compared with some recent state-of-the-art CNN models. The achieved outcomes prove that the suggested model can help radiologists to detect pneumonia-related diseases and improve the diagnostic efficiency compared to the existing diagnosis models.