Detailed morphological analysis of pulmonary structures and tissue, provided by modern CT scanners, is of
utmost importance as in the case of oncological applications both for diagnosis, treatment, and follow-up. In this
case, a patient may go through several tomographic studies throughout a period of time originating volumetric
sets of image data that must be appropriately registered in order to track suspicious radiological findings.
The structures or regions of interest may change their position or shape in CT exams acquired at different
moments, due to postural, physiologic or pathologic changes, so, the exams should be registered before any
follow-up information can be extracted. Postural mismatching throughout time is practically impossible to
avoid being particularly evident when imaging is performed at the limiting spatial resolution. In this paper, we
propose a method for intra-patient registration of pulmonary CT studies, to assist in the management of the
oncological pathology. Our method takes advantage of prior segmentation work. In the first step, the pulmonary
segmentation is performed where trachea and main bronchi are identified. Then, the registration method proceeds
with a longitudinal alignment based on morphological features of the lungs, such as the position of the carina, the
pulmonary areas, the centers of mass and the pulmonary trans-axial principal axis. The final step corresponds to
the trans-axial registration of the corresponding pulmonary masked regions. This is accomplished by a pairwise
sectional registration process driven by an iterative search of the affine transformation parameters leading to
optimal similarity metrics. Results with several cases of intra-patient, intra-modality registration, up to 7 time
points, show that this method provides accurate registration which is needed for quantitative tracking of lesions
and the development of image fusion strategies that may effectively assist the follow-up process.
A segmentation method is a mandatory pre-processing step in many automated or semi-automated analysis tasks such as region identification and densitometric analysis, or even for 3D visualization purposes. In this work we present a fully automated volumetric pulmonary segmentation algorithm based on intensity discrimination and morphologic procedures. Our method first identifies the trachea as well as primary bronchi and then the pulmonary region is identified by applying a threshold and morphologic operations. When both lungs are in contact, additional procedures are performed to obtain two separated lung volumes. To evaluate the performance of the method, we compared contours extracted from 3D lung surfaces with reference contours, using several figures of merit. Results show that the worst case generally occurs at the middle sections of high resolution CT exams, due the presence of aerial and vascular structures. Nevertheless, the average error is inferior to the average error associated with radiologist inter-observer variability, which suggests that our method produces lung contours similar to those drawn by radiologists. The information created by our segmentation algorithm is used by an identification and representation method in pulmonary emphysema that also classifies emphysema according to its severity degree. Two clinically proved thresholds are applied which identify regions with severe emphysema, and with highly severe emphysema. Based on this thresholding strategy, an application for volumetric emphysema assessment was developed offering new display paradigms concerning the visualization of classification results. This framework is easily extendable to accommodate other classifiers namely those related with texture based segmentation as it is often the case with interstitial diseases.
Quantitative evaluation of the performance of segmentation algorithms on medical images is crucial before their clinical use can be considered. We have quantitatively compared the contours obtained by a pulmonary segmentation algorithm to contours manually-drawn by six expert imaiologists on the same set of images, since the ground truth is unknown. Two types of variability (inter-observer and intra-observer) should be taken into account in the performance evaluation of segmentation algorithms and several methods to do it have been proposed. This paper describes the quantitative evaluation of the performance of our segmentation algorithm using several figures of merit, exploratory and multivariate data analysis and non parametric tests, based on the assessment of the inter-observer variability of six expert imagiologists from three different hospitals and the intra-observer variability of two expert imagiologists from the same hospital. As an overall result of this comparison we were able to claim that the consistency and accuracy of our pulmonary segmentation algorithm is adequate for most of the quantitative requirements mentioned by the imagiologists. We also believe that the methodology used to evaluate the performance of our algorithm is general enough to be applicable to many other segmentation problems on medical images.
Bubble emphysema is a disease characterized by the presence of air bubbles within the lungs. With the purpose of identifying pulmonary air bubbles, two alternative methods were developed, using High Resolution Computer Tomography (HRCT) exams. The search volume is confined to the pulmonary volume through a previously developed pulmonary contour detection algorithm. The first detection method follows a slice by slice approach and uses selection criteria based on the Hounsfield levels, dimensions, shape and localization of the bubbles. Candidate regions that do not exhibit axial coherence along at least two sections are excluded. Intermediate sections are interpolated for a more realistic representation of lungs and bubbles. The second detection method, after the pulmonary volume delimitation, follows a fully 3D approach. A global threshold is applied to the entire lung volume returning candidate regions. 3D morphologic operators are used to remove spurious structures and to circumscribe the bubbles.
Bubble representation is accomplished by two alternative methods. The first generates bubble surfaces based on the voxel volumes previously detected; the second method assumes that bubbles are approximately spherical. In order to obtain better 3D representations, fits super-quadrics to bubble volume. The fitting process is based on non-linear least squares optimization method, where a super-quadric is adapted to a regular grid of points defined on each bubble.
All methods were applied to real and semi-synthetical data where artificial and randomly deformed bubbles were embedded in the interior of healthy lungs. Quantitative results regarding bubble geometric features are either similar to <i>a priori</i> known values used in simulation tests, or indicate clinically acceptable dimensions and locations when dealing with real data.
Segmentation of thoracic X-Ray Computed Tomography images is a mandatory pre-processing step in many automated or semi- automated analysis tasks such us region identification, densitometric analysis, or even for 3D visualization purposes when a stack of slices has to be prepared for surface or volume rendering. In this work, we present a fully automated and fast method for pulmonary contour extraction and region identification. Our method combines adaptive intensity discrimination, geometrical feature estimation and morphological processing resulting into a fast and flexible algorithm. A complementary but not less important objective of this work consisted on a quality assessment study of the developed contour detection technique. The automatically extracted contours were statistically compared to manually drawn pulmonary outlines provided by two radiologists. Exploratory data analysis and non-parametric statistical tests were performed on the results obtained using several figures of merit. Results indicate that, besides a strong consistence among all the quality indexes, there is a wider inter-observer variability concerning both radiologists than the variability of our algorithm when compared to each one of the radiologists. As an overall conclusion we claim that the consistence and accuracy of our detection method is more than acceptable for most of the quantitative requirements mentioned by the radiologists.