A New Deep-Learning Method for 3D Lung Cancer Delineation

Lung cancer is one of the most common malignancies worldwide, responsible for a great number of cancer deaths due to late-stage diagnosis, which is typical of the asymptomatic nature of the disease in its early course. It is diagnosed in conjunction with Computed Tomography (CT) to generate high-resolution images that are able to outline tumors. Manual segmentation and analysis of the lung nodules in CT images, on the other hand, is a very time-consuming process and susceptible to subjective error. Recently, deep learning methods, mostly convolutional neural networks (CNN), have had much success in biomedical image segmentation with architectures such as U-Net and 3D U-Net. This thesis will mainly focus on a new approach using a 2D U-Net architecture for automatic segmenting of lung nodules within CT images. This is followed by post-processing continuity analysis, which constructs 3D models of the cancerous regions in a way that clinically relevant tumor delineation can be obtained without more complex and resource consuming models that work with heavier 3D data and features. The model is trained and tested on the NSCLC-radiomics reference dataset of lung cancer segmentation, which contains CT images and their corresponding manual cancer delineation. The performance of the model evaluated using standard metrics to make sure it is clinically viable on the test set shows promising results. In fact, it provided a very good performance on the test set with a high segmentation accuracy (0.9608) of the lung nodules and a Dice Coefficient (DC) of 0.9315. On the entire dataset, for clinical evaluation, the model provides DCs of 0.7997, 0.9123, and 0.9590 for small, medium, and large tumors, respectively. Lastly, DC in 3D reconstruction tests increased from 0.8448 to 0.8850 after post-processing application, hence appropriate volume delineation and cleaning. To our knowledge, this thesis presents a pioneering approach for lung cancer delineation, which for the first time uses the faster and less consuming 2D network to reconstruct data on the 3D level, which provide more clinically relevant information than single-slice imaging. Therefore, also according to the promising results, this work could be considered a forerunner in this line of research, marking a significant advancement in automated medical image analysis.