Assessing the Progression of Pulmonary Emphysema by Advanced Registration

Dutch Technology Foundation, STW LPG.7998

May 2008 - May 2011 

Els Bakker, Denis Shamonin, Marius Staring, Berend Stoel


Chronic Obstructive Pulmonary Disease (COPD) is a lung disease characterized by two components: destruction of lung tissue (emphysema) and inflammation of airways (bronchitis), both of which contribute to airflow limitation and consequent difficulty in breathing. Lung function tests have limited sensitivity and specificity for detecting emphysema, as the bronchitis component affect lung function as well. As a result, there is a great need for alternative markers of emphysema progression, that are more specific and sensitive, so that they can be used in drug evaluation trials and in clinical decision making processes.

In addition to tissue destruction, emphysema is also characterized by increased areas of trapped air and decreased blood perfusion, which together cause a decrease in lung density. Computed Tomography (CT) is used to estimate these changes in lung density and it is now being used as a primary endpoint in drug evaluation trials on emphysema.

COPD occurs in 4 - 10% of the European adult population, while it has been predicted that it will become the most prevalent lung disease and the third leading cause of death in 2020. Since the health and economic burden will increase considerably during the coming decades, the necessity to develop effective treatments for COPD is becoming more and more urgent. However, the evaluation of new drugs has been hampered by a lack of sensitive and specific markers of progression of COPD, and particularly of emphysema. Recently, lung densitometry has been introduced in drug evaluation trials as a primary outcome and alternative parameter to lung function tests. The question arises however, where these changes in lung density occurred, as CT densitometry provides only a global quantification of lung density. This information is crucial for the proper interpretation of the results of an intervention study.

Therefore, we aim to develop methods and software to perform a computerized comparison between baseline and follow-up CT scans by image registration techniques, which needs to be validated extensively. Since one of the main confounding factors is the inspiration level, these image analysis techniques should account for differences in inspiration level. The developed algorithms will be applied directly to image data from a former clinical trial on the natural progression of hereditary emphysema

In first instance, the techniques will be applied to COPD. However, the technology will open up novel applications in other density-related lung diseases, such as asthma, pulmonary edema, acute respiratory distress syndrome and cystic fibrosis, and also in other treatments, such as Lung Volume Reduction Surgery (LVRS).

The project results can be summarized as follows:

  1. Creation of lung densitometry software Pulmo
  2. Local emphysema progression estimation: lobe level
  3. Local emphysema progression estimation: voxel level
  4. Local emphysema progression estimation: visualization

Results: Creation of lung densitometry software Pulmo

Previous software for global lung CT densitometry (PulmoCMS, Medis specials,, frequently used in clinical trials for new drugs for emphysema) was fully rewritten. Rewriting of the software enables easier integration of new techniques and algorithms. The new software QPulmo is made commercially available, and is scheduled to replace PulmoCMS.

Figure 1

Results: Local emphysema progression estimation: lobe level

Densitometry for each lung lobe separately would facilitate a more specific measure for local progression, since emphysema may be confined to one lobe only. Planning of surgical procedures in the treatment of end-stage COPD, such as lung volume reduction by surgery or by placement of endobronchial devices or biodegradable agents, could be made more effective if more localized information (volumetry and/or densitometry) could be provided on the lobes. APPEAR_Lobes

To this end we created an automatic method for lung lobe segmentation, based on enhancement of the dividing structure, i.e. the fissure, combined with an iterative B-spline fitting approach. The latter interpolates in areas missing fissure information and iteratively includes objects most likely to be part of the fissure, which allows detection of the fissure in multiple disconnected parts. In other words, holes in the fissure segmentation are nicely bridged and incomplete fissures are smoothly extended towards the lung boundary. As an additional feature the method allows for effective manual correction by providing extra anchor points, a feature essential for clinical use. Details are given in the conference submission [6].

Fissure detection is based on i) fissure enhancement, ii) removal of vessels, iii) removal of noise, resulting in fissure object candidates, and iv) object splitting at junctions to detach fibrotic tissues, and to split multiple fused fissures. The process is illustrated below, for a right lung, where from left to right the detected objects after step i) till iv) are shown.

Methods were evaluated on thin slice CT data (0.7 x 0.7 x 0.5 mm) from the GLUCOLD study (23 patients suffering from COPD, GOLD stage II and III). A manual ground truth was created using ITK-SNAP (, which was inspected and approved by a pulmonologist. Results were evaluated by comparing the automatic results to the ground truth using the Dice overlap measure. This resulted in an overlap of the fissure (defined 3 voxels wide) with ground truth of 0.65, 0.54 and 0.44 for the three main fissures, which compares to complete lobe overlaps of 0.99, 0.98, 0.98, 0.97 and 0.87 for the five main lobes. More results are given in [6]. This shows promise for lobe segmentation on data of patients with moderate to severe COPD.

Results: Local emphysema progression estimation: voxel level

To tackle the problem of local emphysema progression estimation over time, we proposed to use nonrigid registration of follow-up CT scans, followed by post-processing. Four models were introduced, i.e. four different post-processing steps

  1. Simple subtraction of the registered data. 
  2. Subtraction of the data while compensating for lung volume differences (lung behaves as a dry sponge).
  3. Extension of model B such that the relation between volume and density is not fixed, but estimated globally or locally, based on a third CT scan.
    1. Global estimation of this relation
    2. Local estimation of this relation

Model A is a naïve approach and was shown to fail for cases with volume differences between baseline and follow-up. Lung volume should be taken into account, since it has a major influence on lung density. The spatial Jacobian of the geometrical transformation resulting from image registration can be used to estimate local volume change. Model B assumes that the log volume - log density slope is globally fixed to -1. Empirical evidence (see Stoel et al., "Volume correction in computed tomography densitometry for follow-up studies on pulmonary emphysema," PATS 5, 919 – 924, 2008), together with the observation that lung mass is not constant over the breathing cycle due to e.g. variability in blood perfusion, suggest to adapt the sponge model B. Therefore, model C locally estimates the relation between volume and density. Technical details can be found in the SPIE contribution cited below [1].

The models were evaluated on data of phantoms mimicking lung tissue and on patient data. The results show that model B and C outperform A, as expected. Model B and C performed similarly, model C having a smaller error in correctly estimating the progression of emphysema. See [1] for the first results of validation. A paper with the full validation is in preparation.

Results: Local emphysema progression estimation: visualization

The progression estimation technique was applied to clinical data from a former drug evaluation trial from Roche (REPAIR trial). For this we developed a visualization program to display the progression maps as realistic as possible, while preserving an anatomical reference for navigation. Therefore, we chose to present the maps as a color-coded overlay on top of perpendicular cross-sections in axial, coronal and sagittal direction:

In the middle two columns, the original baseline CT scan and matched follow-up scan is presented, respectively. In the left column, the progression map is displayed calculated from the local slope method, and in the fourth column, the progression map according to the sponge model is displayed. In these progression maps, a green overlay indicates a decrease in density of more than -45 HU and the yellow overlay indicates an increase in density of more than 45 HU. Any change smaller than 45HU is considered not significant, and is not displayed in an overlay. Instead the original CT intensity values are displayed. The red areas indicate lung vessels and areas excluded from the segmentation by Pulmo. In the right column, two images of 3-D volume rendering are displayed to indicate decrease and increase in density, separately, from the sponge model. The graph gives the distribution of local progression values (sponge model). In the results section, below the graph, the median, quartiles and quartile range is shown for the local slope and sponge model. And finally, the local values at the crosshairs in the left most four columns is presented in a table.


In summary, we created software for global lung CT densitometry. Additionally, two methods were devised to enable more local assessment of emphysema progression: one method is lobe-based, for which a lobe segmentation algorithm was made, and the other is voxel-based, for which methods using image registration were proposed. All methods have been validated using phantom and patient data and show that local estimation of the progression of emphysema, which is a slowly progressing disorder, is possible!

Publications from this project

  1. M. Staring, M.E. Bakker, D.P. Shamonin, J. Stolk, J.H.C. Reiber and B.C. Stoel, "Towards Local Estimation of Emphysema Progression Using Image Registration," SPIE Medical Imaging: Image Processing, Proceedings of SPIE, vol. 7259, pp. 72590O, Orlando, Florida, USA, February 2009.
  2. M. Staring, J. Stolk, M.E. Bakker, D.P. Shamonin, J.H.C. Reiber and B.C. Stoel, "Local progression estimation of pulmonary emphysema using image matching," European Respiratory Society Annual Congress, Vienna, Austria, September 2009.
  3. C. Xiao, M. Staring, D.P. Shamonin, J.H.C. Reiber, J. Stolk and B.C. Stoel, "A Strain Energy Filter for 3D Vessel Enhancement," Medical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science, vol. 6363, pp. 367-374, September 2010
  4. M. Staring, S. Klein, J.H.C. Reiber, W.J. Niessen and B.C. Stoel, "Pulmonary Image Registration With elastix Using a Standard Intensity-Based Algorithm," Medical Image Analysis for the Clinic: A Grand Challenge, Workshop Proceedings of MICCAI, EMPIRE10 challenge, 2010.
  5. C. Xiao, M. Staring, D.P. Shamonin, J.H.C. Reiber, J. Stolk and B.C. Stoel, "A Strain Energy Density Method for 3D Vessel Enhancement with Application to Pulmonary CT Images," Medical Image Analysis, vol. 15, no. 1, pp. 112 - 124, February 2011.
  6. D.P. Shamonin, M. Staring, M.E. Bakker, C. Xiao, J. Stolk, J.H.C. Reiber and B.C. Stoel, "Automatic Lobe Segmentation of Emphysema Patients using Iterative B-spline Fitting," submitted to a conference.
  7. K. Murphy, B. van Ginneken, J.M. Reinhardt, S. Kabus, K. Ding, X. Deng, et al., "Evaluation of Registration Methods on Thoracic CT: The EMPIRE10 Challenge," IEEE Transactions on Medical Imaging, 2011, in press.


For more information on this project, please contact Marius Staring: