Convolutional Neural Networks for Automated Detection of Lung Nodules from CT scans

The thesis focuses on training a Deep Convolutional Neural Network (CNN) to recognize lung nodules, using Computed Tomography (CT) scans as training data. An image preprocessing algorithm has been implemented to improve the usability of the images by the Neural Network. A dataset is then created from a few CT scans provided by the Rizzoli Orthopaedics Institute (IOR) for the purpose of training and testing. The starting images are carefully divided and augmented to increase the size of the starting dataset, which originally consists of only 23 images. Subsequently, a portion of these images is used to train the CNN. To improve the performance of the CNN, Transfer Learning techniques are utilized, and four pre-trained networks are used: VGG-16, VGG-19, Inception-V3, and ResNet50. In terms of classification performance, the trained CNNs achieves satisfactory results, with an accuracy value of 95.1%, a precision value of 91.8%, and a recall value of 100% considering the limited amount of starting data. Additionally, the publicly available LUNA (LUng Nodule Analysis) dataset is utilized to conduct further testing and verify the efficacy of the results obtained from the dataset created using the images provided by IOR. This work highlights the potential of deep learning techniques in the field of medical imaging and presents a promising approach for the automated detection of lung nodules.