Pulmonary Lung Classification and Cancer Prediction using Fuzzy Logic and Deep Learning

Abstract

Lung Cancer causes the majority of deaths worldwide and therefore early detection of lung cancer makes the fair chance of patient to survive. Performing chest radiography is the first investigative step towards lung cancer detection. Computed Tomography is state of art technique for diagnosis through imaging. Usually Computed Tomography (CT) scans produces more data from various angles allow a user to see inside the object without any cut. Early Detection of lung cancer for a radiologist is to observe lung pathology usually to observe the presence of nodule from CT scan images. A computer-based process to early predict lung cancer usually reduce the workload of a radiologist and also able to bring better survival rate in no of lung cancer patients. From CT scan image segmenting lung parenchyma is the first and most important steps. Without any loss of a lung, the portion is usually major as well as challenging step. The requirement of segmentation is reduced the computational overload of machines for processing.Various segmentation techniques are still studied in research, as no segmentation techniques give satisfactory results due to variation in lung structures that makes challenging to machines to the segment with better accuracy. Fuzzy logic is been well-known technique and with the wide scope of applications as it uses human behavior analogy. Our aim is to bring human behavior analogy from fuzzy logic into machines to take better decisions. We are using CT scan image in Digital Imaging and Communications in Medicine (DICOM) format, DICOM is standard for storing and transmitting medical images.CT scans for a single patient was generated from 2D lung slices forming a 3D cube. For segmentation, a 2D slice is used from which lung portion extraction was our vital task. For segmentation, we are using fuzzy connectedness method, one of the region based method. Fuzzy connectedness takes global connectivity into consideration and its composed of affinity as local connectivity. Affinity is composed of homogeneity and object features which make decision-making parameter for making the connection of pixels. After doing segmentation by object feature and using morphological operations we are able to extract lung parenchyma better than thresholding method. Deep learning is state of art field of machine learning giving comparative better performance than older machine learning algorithms. By observing the current advantage over previous image classification techniques we go forward to predict lung cancer by using Convolution neural network model of the deep neural network. We used already generated 3D CNN architecture. We had given input of 15 patients to 3D CNN generated model. In our model, we reduced the architecture from orignal defined model to reduce the computational time from 377 seconds to 172 seconds for each epoch. We had done some modifications by initializing weights using the truncated normal method and generated results by Accuracy and Loss. We are able to train the model with a decrease in loss from 9 to 6.5 and accuracy increase from 45% to 65% for 15 patients, validation for our model is still in progress.