Transfer Learning Ensemble Model for Lung Nodule Classification

At present Lung Cancer is the second largest death threat to the world following heart attack. Lung cancer develops a tumor in the lung and the tumor spreads with time if it is malignant. This tumor can be detected in computed tomography(CT) Scan imagery as it shows spots in the lung region. Sometimes it is difficult to detect these spots with human eye as we have some limitation as to how much we can differentiate finer details and most of the time it is detected at later stages. Using Convolution Neural Network we can however detect much finer spots in the lung region and thus can go for further examination. CT-scan images has been used in this paper. The preprocessing operations like median filter for noise reduction and Otsu's method for thresholding has been used to make the images more prudent for classification. Three pretrained Convolutional Neural Network(CNN) models VGG16, ResNet50 and InceptionV3 have been used with modifications. Classification layers of these models have been removed and reconstructed to train on the specific dataset. Images from the preprocessing operations will be fed into the convolutional layers of these models and then output from these will be concatenated and will act as initial layer for the newly constructed dense layer having output nodes of three classes normal, benign and malignant. On the testing dataset this model provides an excellent accuracy of 98%.