| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Paper |
1 Philips Research Laboratories Hamburg, Germany, 2 Department of Radiology, Charité Hospital, Humboldt University Berlin, Germany, 3 Philips Medical Systems CT, Haifa, Israel, 4 Philips Medical Systems CT Clinical Science, Cleveland, OH, USA, 5 Philips Medical Systems Medical IT, Best, The Netherlands and 6 University Hospitals of Cleveland, Ohio, USA
 |
Abstract |
|---|
 Top Abstract IntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
|
Introduction |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
Early detection of lung nodules is crucial both for close observation or biopsy to differentiate between benign or malignant nodules and for timely therapy. Among the most common methods to detect pulmonary nodules are chest radiography and CT. Fibre optic bronchoscopy is also used but has limited value for finding nodules other than those in the larger airways. CT offers better contrast than chest radiography between nodule and background, with no overlapping structures, and several studies have shown that CT can detect smaller, earlier stage nodules with a higher sensitivity than chest radiography [4].
Recently, CT technology has undergone a major evolution with the introduction of multislice technology. With multislice CT (MSCT), a full lung, thin slice (<1 mm) scan can be performed within a single breath-hold. It is hoped that with the high resolution CT data available from MSCT scanners, cancerous nodules can be recognised while still small and in an early stage of lung cancer. Many researchers assume that this downstaging effect by early detection of lung cancer will ultimately improve the survival rate [5, 6].
Moreover, it is hoped that lung cancer screening of high-risk patient groups may significantly increase the rate of lung cancer cases that are diagnosed before the cancer has metastasised. These propositions will be investigated during a large-scale randomised 9-year trial conducted by the US National Cancer Institute (NCI): the currently starting National Lung Screening Trial (NLST) will enrol 50 000 current or former smokers and will take place at a total of 30 sites throughout the USA [http://www.cancer.gov/newscenter/NLSTQA].
Two main areas must be distinguished where computer assistance can be used. First, there is the computer-aided detection (CAD) of pulmonary nodules as such, be they malignant or benign, calcified, solid or subsolid. The results of CAD are markers that draw the attention of the reader to locations of suspicious anomalies. In the context of CAD, a marker is considered as a true positive even if it points at a benign or calcified nodule; false positive markings are then those that do not point at nodules at all (but at scars, bronchial wall thickenings, motion artefacts, vessel bifurcations, etc.). The outcome of CAD is not a yes/no decision, but rather markings at certain locations; therefore the term "true negative" is not defined and a normalised specificity cannot be given. Instead, the performance of CAD is usually given as a sensitivity (detection rate) and a false positive rate (false positive markings per CT study).
The second area is computer-aided diagnosis (often abbreviated as CADx) and aims at the differential diagnosis between malignant and benign pulmonary nodules. Only a fraction of the pulmonary nodules are malignant carcinomas from lung cancer or metastases from cancers in other organs. It is well known that the fraction of nodules that are actually malignant decreases when smaller and smaller nodules are considered [5, 6] as they become detectable by the still increasing resolution of MSCT scanners.
A number of different clinical approaches are aimed at differential diagnosis: biopsy, observation of possible growth by follow-up examinations [7], appraisal of morphological features (such as spiculated or smooth margins), measurement of contrast enhancement in a dynamic CT series [8, 9], and use of additional modalities such as positron emission tomography (PET), etc. For all these clinical approaches, software modules for computer-aided quantification can be used, not only to speed up the workflow, but also to make the necessary measurements themselves more accurate, less prone to error and more repeatable. Going beyond pure quantification, CADx software can then compare the measured values in a multidimensional feature space to known benign and malignant example populations, and retrieve similar cases with known diagnosis from a database or give a likelihood estimation for malignancy [10]. In the context of CADx, false positives are then nodules that are judged to be malignant while really being benign nodules. CADx faces a classical yes/no decision problem, or at least an outcome range that can be normalised to a 0 to 1 interval, so that its performance can be rated by normalised sensitivity and specificity.
Among the various needs for quantification of lung nodules, volume measurement (volumetry) of a detected nodule is the most immediate (for reporting) and the most basic (for detection of possible growth in a follow-up examination).
As a software technology, CAD and volumetry aims at three main objectives.
From a technical point of view, a number of separate technical tasks can assist the reading radiologist, the underlying principles, problems and results of which we will address in this paper:
This paper resulted from a presentation at the British Institute of Radiology on current developments in CAD of pulmonary nodules. The aim of this paper is to show that computer-aided detection and quantification of pulmonary nodules have reached a maturity level such as to significantly contribute to quality of clinical diagnosis.
|
Computer-aided detection (CAD) |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
). This may be even the more so, as modern multislice scanners can produce up to 1000 slices for a thoracic CT examination, where all the small vessels showing up in the isotropic submillimetre resolution have to be checked for their three-dimensional (3D) connectivity to rule out the presence of possible nodules. Therefore, computer assistance for detecting lung nodules in CT data sets is a straightforward concept and has been suggested and investigated as early as 1989 [11–13]. The underlying idea is not that the diagnosis is delegated to a machine, but rather that a machine algorithm acts as a support to the radiologist and points out locations of suspicious objects, so that the overall sensitivity (detection rate) is raised. This could be particularly important in screening situations, with a massive reviewing load of CT studies.
|
ge;5 mm. The principal problem of computerised nodule detection in thick slices of 5–10 mm is that the opacity of nodules that are smaller than the slice thickness is reduced to subsolid appearance by the partial volume effect. Therefore it is impossible to set a certain Hounsfield value as a threshold for potential nodules. Rather, many different thresholds have to be tried [18], and false positives cannot be rejected simply owing to a low Hounsfield value. Another problem with thick CT slices is that a reasoning mechanism is required to recognise constellations where a small nodule is overshadowed by a vessel of similar size within the 10 mm projection [22]. Also, thin vessels may appear disconnected when running obliquely through the slice images and thus be mistaken for nodules.
With the advent of MSCT scanners it is now possible to obtain thin slice CT image data (e.g. 1 mm slice thickness) for the whole thorax within a single breath-hold (Figure 2). For example, at 60 mm coverage per second for a 40-slice scanner (detector coverage 40 x 0.625 mm=25 mm, at pitch 1, no overlap, 0.42 rotations per second), it takes less than 10 s to cover the entire lung area. Owing to the short recording time, multislice high resolution CT data have relatively low motion artefacts, and with a slice thickness 
1 mm allow the detection of very small nodules.
|
The obvious advantage of HRCT data is that with voxel spacings as fine as 0.5 x 0.5 x 0.5 mm3 the partial volume effect vanishes for nodules of size >1 mm (at least for the central voxel). Thus, the Hounsfield values can be evaluated as absolute attenuation values, and the fully isotropic 3D shape information of potential nodules can be utilised by CAD algorithms to resolve ambiguities between pulmonary nodules and vessels.
The CAD algorithm we have suggested in reference [30] works on pure geometric reasoning. It assumes that any blob-like solid structure (which may on one side be attached to a plane-like structure, the lung wall) may be a candidate for a pulmonary nodule if all connecting vessels are significantly smaller in diameter than the central blob-like structure.
For efficiency reasons, these criteria are first checked in the 2D slice images. From each of the resulting seed points, a 3D region-growing procedure is then started to distinguish nodules from vessels that continue with approximately constant diameter, as for nodules it is expected that any connecting vasculature will be decreasing rapidly in diameter. To allow for the wide range of anatomical variability of pulmonary nodules, we use a combination of several geometric descriptors to decide whether a candidate will finally be judged to be a highly likely nodule candidate to be marked by the CAD system.
CAD results on clinical data: detection rate and false positive rate
A pilot study of a lung nodule CAD system (Philips Research, Hamburg, Germany) was conducted on images from the Radiology Department at the Charité University Hospital in Berlin [30]. The images were acquired in the years 2000–2001 by a 4-slice scanner over the entire thorax at 120 kV and 100 mAs.
Each thorax data set comprised 300–500 slice images with 512 x 512 pixels. The x-y-resolution (inslice) varied between 0.5 mm and 1.0 mm, and the z-resolution (reconstruction interval) varied between 0.5 mm and 1.0 mm, with a slice thickness of 1 mm (slightly overlapping slices). In addition to the radiologist, the thoracic data sets were read by the CAD software (computation time was approximately 1 min for all 300–500 slices per thorax). The CAD software marks possible nodules on the axial slice images (Figure 3) and gives an overview of all found nodule candidates for inspection by the radiologist (Figures 4 and 5).
|
|
|
). All nodule candidates were above –400 Hounsfield units (HU), including pleural nodules (Figure 7). Comparison was carried out per nodule for calculation of true and false positives, false negatives, and with respect to nodule size.
|
|
|
However, we would not necessarily expect such a high detection rate at an equally low false positive rate in all clinical settings. Examinations may be taken with low dose or ultra low dose imaging protocols, patient compliance to not breath or move may be less, and moreover the delineations and regularity of the nodule population in question may vary considerably.
CAD performance comparison
Experience with clinical studies has shown that the measured detection rates achieved by CAD systems as well as by radiologists themselves clearly depend on the number of co-reading radiologists: the more co-readers participating, the more suspicious lesions will inevitably be found, and thus the individual sensitivity of each participating radiologist and CAD system will decrease. But even though the absolute sensitivity figures have to be appreciated with care, all clinical studies have agreed that a significant number of nodules have been detected by the additional CAD software alone, while being overlooked by all co-reading radiologists [30–33]. Therefore, CAD can be seen as a strong quality assurance tool.
Up to now it is hard to compare the performance of the different CAD algorithms, as they have all been tested in different settings, on different patient data sets, and acquired with different scanners and imaging protocols.
To allow a more objective comparison of CAD performance, the NCI has formed a consortium to build up a lung image database with consensus-based diagnostic findings [http://www3.cancer.gov/bip/lidc_comm.htm], which could then be used to validate and improve CAD software [34].
Typical false positive findings
In medical terminology, benign pulmonary nodules that do not require action are sometimes called false positives. However, in the context of CAD, the term false positive means locations that are indicated by the CAD algorithm but are not nodules at all. Typical false positive markings are:
Since one of the major sources for false positive nodule markings is cardiac motion artefacts, it is clear that the false positive rate can be significantly reduced by using electrocardiography (ECG)-gated CT studies. In a small pilot study at the Cleveland University Hospitals, we have obtained ECG-gated CT scans from a Philips 16-slice scanner. To find the optimal point in the cardiac cycle, the image volumes were reconstructed at 0%, 25%, 50%, 75% and 88% of the R–R cycle. These scans were then subjected to a lung nodule CAD module [30] and the results were compared with a reference reading of two radiologists. The false positive rate varied between 3.5 false positives per patient in the 88% phase and 6 false positives per patient in the 25% phase. This indicates that performance of CAD can be significantly improved by using ECG-gated CT studies reconstructed in the late diastole. However, the additional burden of ECG gating in terms of costs, examination time and exposure dose can probably not be tolerated in all clinical circumstances, e.g. in lung cancer screening.
In summary, utilisation of HRCT data from multislice scanners for computer-aided nodule detection holds the promise of increasing sensitivity while minimising false positives in such a way as to achieve a breakthrough for CAD in clinical practice.
|
Computer-aided volumetry of pulmonary nodules |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
The aim is to detect malignancy at as early a stage as possible on the one hand, but to avoid unnecessary biopsies on the other hand.
The advent of thin slice data from multi-array CT scanners has enabled fully automated 3D segmentation and volume measurement of lung nodules.
3D volumetry promises a better sensitivity to growth than manually-guided diameter measurements in a single image slice, since an actual doubling of the nodule volume means a diameter increase of a factor of only 3
2 or 26%, which for small nodules might easily be overlooked in the measurement error range.
Slice-wise manual delineation of the nodules through several adjacent slices is a tedious task and is operator-dependent. To improve workflow and reproducibility, algorithms for fully automated 3D computer-aided segmentation and volume measurement of nodules in CT data sets have been proposed [24, 36–39].
Measuring the 3D nodule volume falls into two different sub-problems.
). Therefore, an automated algorithm has to make consistent decisions regarding where to cut off the tumour from the surrounding vasculature. These cut-off decisions will always be somewhat arbitrary, but have to be as consistent as possible in order to allow comparisons between different nodules, and even more important between the same nodule as imaged in the baseline and a later follow-up CT scan. This is not a trivial requirement, as not only the volume but also the morphology of a nodule may have changed during a given time interval (e.g. by angiogenesis). Computer-aided volumetry gives rise to a number of questions that we want to address in this section.
The choice of the Hounsfield threshold above which to account for the nodule volume is a crucial parameter, as the number of visible attached vessels increases with decreasing Hounsfield threshold (see Figure 9). The slice thickness also plays a role: with decreasing slice thickness of the CT image data, more and more surrounding vasculature connecting to the nodules becomes visible that has to be cut off by the segmentation algorithm (see Figure 10). In thick slice data, the connecting vessels can be cut off by raising the Hounsfield threshold [36]. This implicitly utilises the partial volume effect, which assigns lower Hounsfield unit values to thin vessels. This principle does not work in the isotropic resolution of thin slice CT, and several algorithms to cope with this problem have been proposed [24, 36–39]. A method to determine the Hounsfield threshold that yields the strongest gradient between nodule and surrounding parenchyma was proposed by Wiemker and Zwartkruis [40].
|
|
). The region growing first grows into all areas that are spherical or lobular around the current seed point. Then, from all boundary points of the current region expansion, a further seed point is automatically chosen that is closest to the centre of the surrounding mass, and from here the region growing continues into all areas that are spherical or lobular around the current seed point. This iterative scheme ensures that the region growth first expands into all "roundish" parts of the nodule, and only later also grows into the attached tubular vessel structures and possibly into adjacent flat parts of the lung wall (see Figure 11). The growth process stops only after it has grown far from the original start seed point. Then, a retrospective decision is taken regarding where the optimal cut-off point during the growth process would have been. We propose to choose the cut-off point such that the structures that are attached to the nodule are cut through at their thinnest connection, or in the case of an attached lung wall, at the strongest inflection. As an objective function we therefore propose the surface integral over the distance values of the 2D surface of the current 3D region expansion, where the distance value for each point is defined as the distance between this point and the closest outside nodule point (the surrounding "dark" parenchyma tissue).
|
For the proposed nodule segmentation algorithm, we have studied the effect of slice thickness on the measured nodule volume in order to determine the consistency and comparability of nodule volumes measured during follow-up examinations [39]. Considering the delicate vasculature network around the pulmonary nodules, the robustness of computer-aided nodule segmentation algorithms cannot be validated on data from CT phantoms with smooth boundaries only.
Analysis of 41 nodules between 150 mm3 and 350 mm3 (6.6–8.7 mm effective diameter) showed that the connectivity of the nodules to the surrounding vasculature clearly increases with thinner slice thickness, so the segmentation algorithm has to cut off more and more thin vessels. On the other hand, in the finest slice spacing there is the weakest dependency of the nodule volume on the Hounsfield threshold (owing to the diminishing partial volume effect). The cut-off decisions for surrounding vessels to be made by the algorithm are different for each Hounsfield threshold, as more or less vessels are visible above a given threshold. Nevertheless, we found that the segmentation algorithm yielded consistent volume estimations with varying slice thicknesses using a threshold of –400 HU. The average variation in volume estimation stayed within 3%. This confirms that the proposed segmentation algorithm yields consistent decisions regarding where to cut off the attached vessels. The segmentation algorithm is shown to be robust enough to compare nodule volumes between follow-up CT studies even of different slice thickness.
In another study at the Hematology–Oncology Clinic of Little Rock, Arkansas, we have compared manual segmentation by two radiologists to a computer-assisted automatic segmentation (Lung Nodule Assessment and Comparison Package; Philips Medical Systems, Best, The Netherlands) for 33 nodules, varying in size from 3.3 mm to 30.0 mm effective diameter. The nodules were classified as difficult or isolated based on whether they were attached to vessels, lung wall, or close to other nodules or scar tissue. Several types of nodules such as solid, spiculated and calcified were included in the study. Using identical image window, level and image zoom, both radiologists drew contours along the boundaries of nodules in every axial image in which the nodule is present. All nodule images were acquired at 16 x 0.75 mm collimation and reviewed at 1 mm slice width and 1 mm increment. The total volume from manual segmentation was calculated by summation of the individual slice-wise contour areas. For automatic segmentation, a seed point was placed on the nodule in only one axial slice.
Comparisons were made of manual measurements between the two radiologists, manual vs automated volume segmentation, and time for segmentation using automated and manual methods. The correlation between the manual measurements by two radiologists was 0.987. There was also a strong correlation between automated and manual segmentation, with the correlation coefficient equal to 0.972 and 0.986 for the two radiologists, respectively. Thus the agreement between manual and computer-aided volumetry proved to be equally good as the agreement between the two human readers. The automated measurements required minimal user interaction, with average volume estimation time per nodule of a few seconds. Manual measurements on the other hand took an average of 5 min per nodule.
Another question is how robust computer-aided nodule volumetry is against varying radiation dose. Hay et al [41] have studied the effect of using low dose imaging protocols for lung nodule volumetry. In their phantom study with 34 different artificial spheres and cylinders of different sizes, they found the segmentation accuracy to be better than 5% both for standard dose and low dose CT scans.
In summary, we conclude that for solid nodules, computer-assisted volumetry is accurate, robust, repeatable and consistent. For subsolid nodules (ground-glass opacities) however, definitions of meaningful characterisation still have to be agreed on.
|
Follow-up registration and matching of nodules |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
). Moreover, a list of matching pairs is automatically compiled for all nodules found in the two data sets (detected both manually and/or by CAD).
|
The automatic nodule matching approach described by Blaffert and Wiemker [42] starts out by segmenting the lungs out of the overall thoracic CT data volume of the current and former study. The two lung volume images are registered (aligned) using an affine coordinate transformation. The similarity between the former and current lung volume is tested by virtue of the cross-correlation of both data sets. The coordinate transformation is then varied until optimal cross-correlation is reached. By using Gauss–Newton optimisation methods in variation of the coordinate transform, and by employing a coarse to fine resolution multiscale approach, the optimal alignment of the two lung volumes can be reached in typically 5 s.
In general, an affine coordinate transformation is not expected to always suffice for the alignment of the same lung between a former and current CT study, since the possibly different respiratory state and patient pose on the CT table may cause elastic deformations. However, elastic image registration is too time consuming at the current state of the art, and the experimental results indicate that affine image registration comes close enough to allow correct matching of the nodule positions in the former and the current study: after alignment of the lungs in the CT study pair, the mean distance between the location of a nodule in the former and the current study was 4 mm [42]. 124 of 125 nodules were correctly matched even if the CT study pairs were recorded with different field of view, slice thickness or slice spacing [43].
Betke et al [44] use a different algorithm and test set and report a mean distance of 8 mm between the location of a nodule after automated image registration, and 56 of 58 nodule pairs correctly matched. Another equally favourable result was given for an algorithm of Shen et al [45].
|
Conclusion |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
As more and smaller lung nodules become detectable in thin slice CT, automated and reproducible computer-aided 3D volumetry of lung nodules becomes important to speed up workflow, to compute growth rates between follow-up examinations and thus to avoid unnecessary biopsies.
With ever refining spatial, dynamic and temporal resolution of CT scanners, more applications of computer-aided detection and quantification are likely to emerge.
|
Acknowledgments |
|---|
|
References |
|---|
|
TopAbstractIntroductionComputer-aided detection (CAD)Computer-aided volumetry of...Follow-up registration and...ConclusionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  A. Marchiano, E. Calabro, E. Civelli, G. Di Tolla, L. F. Frigerio, C. Morosi, F. Tafaro, E. Ferri, N. Sverzellati, T. Camerini, et al. Pulmonary Nodules: Volume Repeatability at Multidetector CT Lung Cancer Screening Radiology, June 1, 2009; 251(3): 919 - 925. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. A. Gavrielides, L. M. Kinnard, K. J. Myers, and N. Petrick Noncalcified Lung Nodules: Volumetric Assessment with Thoracic CT Radiology, April 1, 2009; 251(1): 26 - 37. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 G. Veronesi, M. Bellomi, P. Scanagatta, L. Preda, C. Rampinelli, J. Guarize, G. Pelosi, P. Maisonneuve, F. Leo, P. Solli, et al. Difficulties encountered managing nodules detected during a computed tomography lung cancer screening program J. Thorac. Cardiovasc. Surg., September 1, 2008; 136(3): 611 - 617. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 P. C. Strohm, D. Kubosch, T. A. Bley, C. M. Sprecher, N. P. Sudkamp, and S. Milz Detection of Bone Graft Failure in Lumbar Spondylodesis: Spatial Resolution with High-Resolution Peripheral Quantitative CT Am. J. Roentgenol., May 1, 2008; 190(5): 1255 - 1259. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| BJR | DMFR | IMAGING | ALL BIR JOURNALS |
