LUNGS TUMOR CLASSIFICATION WITH CT SCAN IMAGES USING CONVOLUTIONAL NEURAL NETWORK

Abstract

Lungs cancer is one of important lung disease characterized by the unregulated separation of living organisms into cells. It expands to develop tumors and raises conflict with lung functioning by preventing the lungs from doing their tasks. Hence, the most effective means to minimize lung cancer mortality are early detection and accurate therapy of lung tumors for patient survival. Computed Tomography (CT) for lung cancer screening with images has thus begun to be commonly adopted across the world. Analysis of those CT SCAN images, though, is a massive burden for the oncologist. So, Intelligence computer-aided-diagnostic (CAD) systems are also built to detect the cancer at an earliest stage that assist radiologists and a pulmonary medical expert in making decisions regarding the detection and treatment of their patients. In this study, a new automated approach has been proposed that uses a 2-dimensional convolutional neural network (CNN) model to detect the tumor by Lung CT SCAN images and also classifies them into two types of labels i.e., normal lungs and tumor infected lungs. Various pre-processing techniques are used in proposed system to enhance the quality of the CT SCAN images and a CNN model is created for classification. Dataset of 4,337 Lung CT SCAN images have been collected from the Nishtar Hospital and Nishtar University Multan, Pakistan with 1,028 images of healthy lungs and 3,345 images of tumor infected i.e., Malignant or Benign and the dataset is used in training and validation. Our proposed design network randomly selects both normal and tumor class images from the dataset to give the best result for training. This study used 8:2 ratio of data, i.e., 80% of the data for the training set and 20% of the data for validation or testing. In our study, we differentiate these two classes (Normal, Tumor (Benign, Malignant)) through a softmax linear classifier. The final classification layer predicts the final validation accuracy of the trained network. In the end, the prediction of the labels is computed and based on predicated and true labels of the images, final accuracy is calculated. We benefit ourselves with 31 layers of CNN design with 5 number epochs for each iteration, which is the cycle operating in the neural network system and the learning rate set to 0.01. To validate the results of our model, we used the Alex Net network by altering its two layers. The Classification Accuracy that our research acquired is 98.74%.