A 3D α-hull approach to juxta-pleural nodule detection in chest Computed Tomography images

The aim of this work is therefore to contribute to the development of a CAD system for the automatic detection of juxta-pleural nodules in CT scans.
The basic steps of the CAD system presented in this work are:
- segmentation of the lung parenchyma, after which juxta-pleural nodules appear as concavities of the lung surface;
- concavity closing in the binary segmented mask with the alpha-hull concavity patching method in order to include in the segmented volume those concavities that may be related to juxta-pleural nodules. By applying the difference operation between the closed and the original mask, a list of nodule candidates is obtained;
- feature (candidate characteristic parameters) extraction and calculation for each nodule candidate;
- nodule candidate classification by using Artificial Neural Networks.
In this work, in order to overcome the limitations imposed by the 2D analysis previously performed by our research group, a 3D approach
is performed. The method proposed consists of stacking 2D single slice binary masks whose concavities have been closed by the multiscale alphahull algorithm in order to obtain a 3D reconstruction of the closed lung. By applying a difference operation with the segmented original volume, we detect a list of concavities (either natural or due to nodules). The study of the performance of the nodule hunting based on alpha-hull for a fixed range of α values is also presented.
This study leads to the assessment of the optimum parameter optimizing detection sensitivity, which identifies all juxta-pleural nodules in the CT data set. In order to reduce the number of false positives (regions reported as healthy by radiologists and incorrectly identified as pathological by the CAD system) a research on the most effective characteristics of the nodules is carried out. Finally, the classification step is performed, preceded by 3D features extraction, analysis and filtering.
The classifier developed in this work, is an Artificial Neural Network. Training and test of the network are performed. By varying the number of neurons in the hidden layer, the structure with the best performance is determined. The results obtained are then compared to other CAD systems performance.