Automatic lung classification from x-ray Images of normal and pneumonia patients

Background: Nowadays AI (artificial intelligent) plays an important role in assisting human especiallyexperts in many aspects. As a branch of AI, deep learning is part of a broader familyof machine learning methods based on artificial neural networks with representation learning.Recently, deep learning algorithm have been used widely in health domain including prognosis ofpatients with chronic obstructive pulmonary disease from chest radiographs. This study focuseson using algorithms to detect pneumonia from chest x-rays by using deep leaning.Method: this is a retrospective applied study in which a binary classification method was developedto classify pneumonia patients and normal people based on their chest x-ray images. Firstly,for data collecting, an online dataset containing ۵۸۶۳ x-ray images categorized to normal andpneumonia were accessed. Given that the images in dataset were collected from different sourcesthey came in different sizes and aspect ratios, a resizing strategy devised with average aspect ratioof all images by means of preprocessing. Images channels also divided by ۲۵۵ for rescaling. Thethird step, a CNN model consisting of five layers of Convolutional layer plus max pooling thatfollowed by flattening layer and connected layer with ۳۲ neurons was designed and utilized. Inthe final step, the model was trained using the sigmoid function. It was also compiled with RMSprop as optimizer and binary Crossentropy loss function.Results: True positive(TP), cases were ۴۲۷, true negative(TN), False positive (FP), and falsenegative(FN) were ۴۲۷, ۱۲۲, ۸, and ۲۲ respectively. The last Epoch accuracy, the last Epochvalidation accuracy, and the last epoch validation loss that model achieve was ۹۶.۶۴%, ۹۴.۹۲%,and ۰.۱۶۶۹, respectively. Sensitivity of the model to diagnose pneumonia cases was ۹۴.۹۸% and۹۴.۷۳% of all the normal cases were recognized true (specificity). f۱-score (harmonic mean ofprecision and recall) of the model was ۹۶.۵۳% as important factors for calculating the accuracy.Discussion and conclusion: The layers used in this model, make it different from other CNNmodels which caused high accuracy. In some models, some important layers like MaxPoolling۲Dand Conv۲D are just used, nut in this model, there are some useful layers such as, Spatial PyramidPooling (SPP) which benefits from no need for a fixed size input. It helps to reach better validationaccuracy and reduce the validation loss. Although it is useful for helping radiologists to diagnosepneumonia faster and more accurately, it is hard to develop and needs strong GPU and CPU. Artificialintelligence in radiology has undergone something of a metamorphosis and has grown asboth a technology and a market. However, like most technology solutions, artificial intelligence(AI) is not perfect. It is also, most importantly, not a replacement for human beings.

کلیدواژه ها:

Deep learning ، pneumonia ، x-ray ، image processing