Detection of Pneumonia from X-RAY Imaging Using Faster RCNN: Analytical Essay

NB: We do not resell your papers. Upon ordering, we do an original paper exclusively for you.

NB: All your data is kept safe from the public.

Abstract

Streptococcus pneumoniae is a bacteria (major one) that causes deadly disease named pneumonia. It majorly affects the lung portion. Alveoli is an air sac present in the lungs where the exchanging of oxygen from the lung to the body and exhaling carbon dioxide (CO2) takes place first. The air sac gets filled with fluids, pus, etc which causes inflammation that leads to difficulty in breathing. Nearly a million adults were suffered from this deadly disease. Most of the developed countries also get panicked due to this disease. The initial step in treating pneumonia is to take an x-ray of the chest region. Only minor symptoms are visible in X-ray which is very difficult to interpret. So, it is done only by experts, not by ordinary physicians. The present world gets automated day by day so this problematic pneumonia detection requires automation. Applying deep learning in pneumonia detection will help to get rid off this problem. The proposed idea is to use a deep learning approach called Faster region convolutional neural network (Faster RCNN) instead of analyzing the whole chest x-ray image only the lung region is analyzed.

Keywords—FASTER RCNN; deep learning; Pneumonia; x-ray

I. Introduction

Pneumonia is a deadly lung disease caused by bacteria, fungi, and viruses. The main part of the lung is alveoli which is responsible for the exchanging oxygen (O2) from lung to body and exhaling carbon dioxide (CO2). This alveoli an air sac gets infected by this pneumonia diseases. The air sac gets filled with fluid and also suffers from inflammation. Rising in body temperature (fever), cold, repetitive cough, difficulty in breathing, and frequent tiredness are indications of pneumonia. The indications along with colored sputum i.e., green, bloody sputum or yellowish confirm pneumonia infection. Sputum color indicates the severity of the disease. Based on NCBI pneumonia comes under two types namely community-acquired pneumonia (disease acquired in the home) and nosocomial pneumonia (disease acquired in the hospital). Treating pneumonia mainly involves taking an x-ray of the chest area further diagnosis may require CT scan, MRI scan, ultrasound of the chest depending upon physician treating techniques. Mostly x-rays were used because it is cost-effective and affordable for all class of people. The main obstacle involved here is interpreting x-ray which is not as easy as normal x-rays taken from non-complicated body parts.

So the interpretation of x-ray requires experts not by all physicians. Already previous works were done on using convolutional neural network but all those were implemented using traditional CNN techniques which require high computational cost and involves analyzing unwanted x-ray portions. The proposed idea is to automate pneumonia detection using a deep learning approach named FASTER RCNN which analyses only the desired portion using region of interest (ROI) avoids unnecessary portions in a chest x-ray.

II. Related works

The deep learning technique was mainly used for object detection, facial recognition related applications. The previous related work was done on CNN (convolutional neural network) using Xception and vgg16 network by Enes AYAN and Halil Murat ÜNVER. A comparative research study was done by them. Parameters like training time, speed, and accuracy were measured in previous work.

But theses CNN techniques were replaced by new methods presently. The previous work was not produced notable results only comparative research was done. Layers used in previous work were increased not concentrated new logic. The dataset was classified as only two classes namely pneumonia affected and not pneumonia affected no deeper analysis was done. Pneumonia x-ray shows only slight variations like opaqueness, inflammation and not normal these variations help in accurate detection. The CNN major limitation is that it does not encode object position and orientation. So it leads to a waste of computational cost because x-ray image contains neck, shoulder portion these were not required for analyzing pneumonia.

III. Materials and methods

A. Dataset and data preprocessing

In this research, the pneumonia x-ray image dataset was collected from the Kaggle platform consists of around 26000 images in DCM format and a CSV file with a bounding box and patient information. This dataset was published under RSNA pneumonia detection challenge in Kaggle platform. Around 4000 images were used for this study and these images were classified as three classes namely 1. Lung opacity 2. Normal 3. Not normal. Figures (Fig.1, Fig.2, Fig.3) for respective categories were mentioned for some images, bounding boxes were drawn manually using the LabelImg tool. The XML file produced by LabelImg tool was converted to CSV file using python xml_to_csv program. All x-ray images were in DCM format and these images were converted to JPG format using DICOM converter software. Dataset images were in 1024 X 1024 dimension and the desired region for each image was available in the Kaggle platform and for some images desired bounding coordinates were drawn manually and those bounding box values were updated in CSV file to train images.

Fig.1. Normal

Fig.2. Not normal Fig.3. Lung opacity

B. Faster RCNN architecture

One of the deep learning techniques is FASTER RCNN (Region convolutional neural network) that this method has an important advantage i.e., uses RPN (Region proposal network). The normal CNN layers (Convolutional, ReLu, Pooling, and Fully connected layer) were involved in this method followed by some additional features that play a major role here. RPN is a major player produces a number of proposals from the feature map which was the output of CNN layers. RPN uses classifier and regressor which tells whether the particular proposal is the desired one or background. The proposals are called anchors. After ROI (region of interest) pooling was applied to produce a fixed-size feature map and then the SoftMax layer and regressor suggest bounding boxes. The pros of this method are it avoids selective search and avoids feeding of 2000 regions for each layer. In this study, a faster RCNN inception v2 coco model were used to train the dataset. Pneumonia analysis mainly insists on the lung portion where only major interpretation takes place so RCNN plays a major role by taking only the desired area or region for analysis. Usually, an x-ray of chest contains neck, shoulder region those all are not required in the analysis. This method is an apt one for analyzing pneumonia by using all parameters efficiently. Figure 4 represents architecture of faster RCNN. In figure 4 four CNN layers were involved namely convolutional, ReLu, pooling, and fully connected layer. All these layers produce a feature by interpreting from the input x-ray image and RPN suggests proposals from the input image and then region of interest pooling applied to make all proposals to a fixed size. After that classifier tells whether the particular part infected with pneumonia or not.

Fig.4.FASTER RCNN ARCHITECTURE

C. Model and device

In this study, a faster RCNN inception v2 coco model were used to train the dataset. TensorFlow library was used along with a faster RCNN pretrained model to train the dataset. TensorFlow has its pros such as easy error debugging, graph notations, easy library management, ease to extend and pipelining support were tremendous. The library was updated regularly so it will be fruitful for developing projects. Google colab a free cloud service that offers free GPU to develop deep learning projects with major library support. The specifications were followed by 1xTesla K80 12GB GDDR5 VRAM with CUDA support. Dataset was trained using Google colab to avoid unnecessary training problems.

D. Experimentation results

In this part, training and techniques were discussed. The dataset was trained using faster RCNN inception v2 coco which is an open-source model and the TensorFlow library was used. All images were in 1024 X 1024 dimensions to avoid low-quality image problems the images were used as it is and no images were compressed and the images used are taken from RSNA pneumonia detection challenge. The threshold value for training was set to around 20000 steps to maintain a decent loss percentage. Training loss and RPN loss were illustrated graphically using tensorboard. To reduce the loss percentage number of iterations were carried out until the loss value comes under 0.099. Faster RCNN requires training time a little more compared with other deep learning techniques. The main problem faced was more training time required. Training images around 4000 were used and test images were around 1000 randomly picked from the dataset. Images were classified into three classes lung opacity, normal and not normal. Lung opacity and not normal indicates pneumonia infection and normal represents free from pneumonia infection. Random images were tested and checked whether it correctly classifies or not. Figure 5 represents lung opacity i.e., pneumonia affected and figure 6 represents normal i.e., free from pneumonia.

Fig.5. This image was classified by Faster RCNN shows that person affected by pneumonia (Lung opacity class).

Fig.6. This image was classified by Faster RCNN shows that person free from pneumonia (Normal class).

IV. Conclusion

This paper provided a solution to automate pneumonia detection from x-ray images using faster RCNN without depending on physician analysis. This work not fully replace the need for physician analysis but helps in analyzing the pneumonia. The results were obtained using the Faster RCNN inception v2 coco model. Future work is based on analyzing pneumonia using x-ray along with person-environment and body physical conditions.

V. References

E. Ayan and H. M. Ünver, ‘Diagnosis of Pneumonia from Chest X-Ray Images Using Deep Learning,’ 2019 Scientific Meeting on Electrical-Electronics & Biomedical Engineering and Computer Science (EBBT), Istanbul, Turkey,2019, pp.1-5.
B. Li, G. Kang, K. Cheng, and N. Zhang, ‘Attention-Guided Convolutional Neural Network for Detecting Pneumonia on Chest X-Rays,’ 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 2019, pp. 4851-4854.
S. R. Islam, S. P. Maity, A. K. Ray and M. Mandal, ‘Automatic Detection of Pneumonia on Compressed Sensing Images using Deep Learning,’ 2019 IEEE Canadian Conference of Electrical and Computer Engineering (CCECE), Edmonton, AB, Canada, 2019, pp. 1-4.
Stephen, Okeke & Sain, Mangal & Maduh, Uchenna & Jeong, Doun. (2019). An Efficient Deep Learning Approach to Pneumonia Classification in Healthcare. Journal of Healthcare Engineering. 2019. 1-7. 10.1155/2019/4180949.
D. Varshni, K. Thakral, L. Agarwal, R. Nijhawan and A. Mittal, ‘Pneumonia Detection Using CNN based Feature Extraction,’ 2019 IEEE International Conference on Electrical, Computer and Communication Technologies (ICECCT), Coimbatore, India, 2019, pp. 1-7.
Jaiswal, Amit & Tiwari, Prayag & Gupta, Deepak & Khanna, Ashish & Rodrigues, Joel. (2019). Identifying Pneumonia in Chest X-Rays: A Deep Learning Approach. Measurement. 145. 10.1016/j.measurement.2019.05.076.

Click Here To Order Now!