Solitary Pulmonary Nodule malignancy classification utilising 3D features and semi-supervised Deep Learning

The volumetric representation of Solitary Pulmonary Nodules (SPN) in Computed Tomography (CT) imaging is mandatory, especially for capturing and analysing deep features and having a complete picture of the morphology, the shape of the volume, its distribution in space, and its relationship with the adjacent tissues. Automated deep feature extraction in three dimensional space is a specialisation area of 3D Convolutional Neural Networks (CNN). The extraction of the most representative features of malignant SPN representations, can be achieved with the assistance of CNNs. To evaluate this methodology, a 3D CNN called 3D-LidcNet, is developed in this study. The Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset is utilised to extract 2124 SPNs represented in sets of 2D slices. By concatenating 16 slices for each SPN, 3D nodule representations are constructed. To increase the learning capabilities of the 3D CNN, data augmentation is applied during training. 3D-LidcNet achieves 90.68% accuracy in distinguishing benign from malignant SPNs, coming from the strongly labelled subsets of the dataset (898 unique SPNs). To make full use of the weakly labelled SPNs, a semi-supervised training algorithm is utilised to progressively expand the training dataset with the most confident predictions of the weakly labelled SPNs. This approach succeeds in classifying 1585 SPNs, with an accuracy of 87.44%. Finally, 3D-LidcNet is trained and tested using the complete dataset (2124 SPNs) to distinguish between benign and malignant nodules, achieving an accuracy of 89.68%.