CAD in clinical trials: Current role and architectural requirements

6 downloads 451 Views 2MB Size Report
architectural requirements to incorporate CAD into clinical trials and illustrate their functionality in therapeutic trials for .... The ability to edit segmentation results is vital since “auto- ... is run and then initially edited by a laboratory technician. The.
Computerized Medical Imaging and Graphics 31 (2007) 332–337

CAD in clinical trials: Current role and architectural requirements Matthew S. Brown ∗ , Michael F. McNitt-Gray, Richard Pais, Sumit K. Shah, Peiyuan Qing, Irene Da Costa, Denise R. Aberle, Jonathan G. Goldin UCLA, Department of Radiological Sciences, United States

Abstract Computer-aided diagnosis (CAD) technology is becoming an important tool to assess treatment response in clinical trials. However, CAD software alone is not sufficient to conduct an imaging-based clinical trial. There are a number of architectural requirements such as image receive (from multiple field sites), a database for storing quantitative measures, and data mining and reporting capabilities. In this paper we describe the architectural requirements to incorporate CAD into clinical trials and illustrate their functionality in therapeutic trials for emphysema. © 2007 Elsevier Ltd. All rights reserved. Keywords: Computer-aided diagnosis; Clinical trials; Emphysema

1. Introduction Computer-aided diagnosis (CAD) technology is being developed for detection, quantitation and change assessment and has potential applications in therapeutic assessment. Computeraided diagnosis systems have focused on lesion detection applications in mammography, colonography and lung nodule detection [1–5]. Systems have shown promising experimental results, and some commercial systems are now available. There have also been significant efforts to develop and test lesion size measurement tools [6–8]. These remain active areas of CAD research to improve detection performance and to validate systems for response assessment. This research is vital if CAD detection systems are to become an integral part of oncology clinical trials, for although they have great potential they are not yet considered essential. However, in the domain of diffuse lung disease, and in particular emphysema, CAD is now playing a central role in clinical trials of new therapies. Emphysema results in the destruction of lung parenchyma. Lung compliance increases and, because of the loss of radial traction on the conducting airways, the airways collapse prematurely during expiration, causing airflow obstruction and the inability to breath. Emphysema is a major cause of morbidity and mortality for an estimated 60 million ∗ Corresponding author at: Thoracic Imaging Research Group, 924 Westwood Blvd, Suite 650, Los Angeles, CA 90024, United States. Tel.: +1 310 794 8969; fax: +1 310 794 8647. E-mail address: [email protected] (M.S. Brown).

0895-6111/$ – see front matter © 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.compmedimag.2007.02.014

patients world-wide. CT imaging can effectively depict emphysema severity (see Fig. 1), and it can be quantitated by density mask [9,10]. The density mask measure is the percentage of voxels below a particular Hounsfield Unit (HU) threshold, which corresponds to air-containing holes in the parenchyma. Thus, CAD software that can segment the lung anatomy and derive quantitative measures such as density mask scores can allow us to assess the severity of emphysema. This information can be used to identify patients that are candidates for treatment and to assess the response of patients post-treatment. However, CAD software alone is not sufficient to conduct an imaging-based clinical trial. There are a number of architectural requirements such as image receive (from multiple field sites), a database for storing quantitative measures, and data mining and reporting capabilities. In this paper we will describe the architectural requirements to incorporate CAD into clinical trials and illustrate their functionality in therapeutic trials for emphysema.

2. Methods The key functional components for CAD in clinical trials are: DICOM image receiver; CAD system for quantitation; CAD review workstation; statistical analysis workstation; and web server for reporting of results. These components are linked through a database management system as shown in Fig. 2. We will illustrate the functionality and interaction between these components in emphysema treatment trials.

M.S. Brown et al. / Computerized Medical Imaging and Graphics 31 (2007) 332–337

Fig. 1. Lung CT image showing regions of bullous emphysema destruction, predominantly in the lower lobes (arrows), with computer-determined involvement percentages.

333

creating additional space in the chest cavity to allow the adjacent (healthier) lung to expand. However, both techniques are highly invasive and carry significant risk of complication. New minimally invasive therapeutic options for emphysema are now currently being tested in clinical trials. One such example is the endobronchial valve for emphysema palliation trial (VENT) conducted by Emphasys Medical Inc. (Redwood City, CA) [11]. A one-way valve (Emphasys Zephyr® Endobronchial Valve) is inserted bronchoscopically into the airway feeding the most severely diseased lobe. This valve only allows air to escape from the lobe, thereby inducing lobar volume reduction and achieving a result similar to lung volume reduction surgery without a major surgery, therefore posing less risk to the patient [12–16]. CAD technology has been central to this trial. A CAD system is being used to segment the lung into its component lobes and to assess the severity of emphysema in each lobe. This enables determination of emphysema heterogeneity (eligibility of the patient for treatment) and the most severely affected lobe into which the valve should be placed.

2.1. Emphysema treatment trials

2.2. Image management

Currently there are limited treatment options available for patients suffering with emphysema. For advanced stages of the disease a lung transplant or lung volume reduction surgery can be considered. The latter technique can only be applied in cases with heterogeneous emphysema, i.e., when one lobe of the lung is emphysematous and the adjacent lobe less severely effected. The more diseased lobe is surgically removed in the hope of

Image transfer and storage capabilities are an essential component of the clinical trials infrastructure. The VENT involves patient data collection at 57 clinical sites in the United States and Europe. These images are transferred by DICOM push or by mail on CD (if network transfer rates are limited). A CAD infrastructure also requires image storage with redundancy for fault-tolerance. In this case we use a pair of mirrored data servers

Fig. 2. Overview of radiology core infrastructure and data management.

334

M.S. Brown et al. / Computerized Medical Imaging and Graphics 31 (2007) 332–337

each with RAID 5. Grid-based approaches are also being developed for image transfer and fault-tolerance [17]. Imaging protocols in clinical trials usually contain multiple series. For example, the VENT protocol involves thick and thinsection CT imaging at two suspended breathold levels: total lung capacity (TLC) and residual volume (RV). These series must be consistently labelled so that they can be organized within the database prior to CAD processing. The labelling is determined at the core laboratory by review of the images and DICOM fields. Image quality control and protocol adherence at clinical sites are vital for reproducible quantitation. For each imaging study received at the core lab, study quality is checked for adherence to protocol and a numerical grade is given. This score and comments are included in an email to the site to provide ongoing feedback and maintain image quality. The scores also provide important monitoring information for sponsors on overall image quality in the trial. To support these requirements a database management system is essential. Our data model is based on the DICOM hierarchy that is augmented as follows. The Image table contains a file path to the location of the actual DICOM files (images) on the data server. There are additional table associated with the Series table that contain fields for labelling specific series and storing image quality scores and comments.

2.3. CAD analysis CAD techniques are used to provide automated image segmentation and quantitation. For the VENT lobar lung segmentation is required. A model-based approach for automatic segmentation of the lung parenchyma in CT has been previously published [18,19]. To subdivide the lungs into lobes a semi-automated approach has been developed. On thick-section images the major and minor inter-lobar fissures are drawn manually on the axial images (see Fig. 3c). This is done manually since the fissural surfaces are thin and difficult to detect due to partial volume averaging (see Fig. 3b). After the fissures are contoured, an automated region growing technique is applied to extract the lobes as separate regions of interest (ROIs). On thin-section images the fissures are more visible, as shown in Fig. 3a, and techniques are being developed that use gradient and 3D continuity to automatically detect the fissural surface [20]. However, in the VENT thick-section imaging is being used to provide the primary measure of emphysema because density mask measurements were originally validated against pathology specimens using 10 mm thick slices [9]. The ability to edit segmentation results is vital since “automated” CAD systems are often inaccurate for clinical trials patients with disease patterns different to those on which the sys-

Fig. 3. Semi-automated lobar segmentation for a CT scan acquired at residual volume (RV). (a) Thin-section CT showing inter-lobar fissures (arrows). (b) Corresponding thick-section image (from same scan) with partial volume averaging leading to low contrast of the fissures. (c) Approximate manually drawn fissures on thick-section. (d) Automated segmentation of left upper, left lower, right middle and right lower lobes (right upper lobe is not visible on this slice).

M.S. Brown et al. / Computerized Medical Imaging and Graphics 31 (2007) 332–337

tem was developed (trained). For example, emphysema distorts the architecture of the lungs and their fissural surfaces meaning that automated systems often fail and must be edited. For this to be practical in clinical trials, the automated algorithm is run and then initially edited by a laboratory technician. The segmentation is then approved by a radiologist. Following segmentation the system then computes the density mask (DM) as the percentage of pixels within each lobe that are below −910 HU (a threshold determined by Ref. [9]). Based on this score, a patient is eligible for treatment if they have heterogeneous emphysema, i.e., different scores in adjacent lobes. The CAD system can then also determine the target lobe as the most severely diseased, i.e., highest DM score. To support these operations the database contains tables for storing segmentation results as ROIs [21], quantitative data and user information. The data model also allows a radiologist to sign-off on a case at which point it becomes locked in the database. The necessary tables are associated with an image series, i.e., with rows in the Series table. 2.4. Statistical analysis and reporting Results for an individual patient are reported in Portable Document Format (PDF) via the web (see Fig. 4). Clinical sites

335

have password-protected access to reports for their patients. This web-based report shows status information for a given patient (whether images have been received and analysed) as well as density mask scores and lobar targeting information. These forms are generated on the fly by querying the database, meaning that the moment a radiologist approves a case at the core lab the report is available online at the clinical site. The database can also be mined to perform quality control (QC) checks and to perform statistical analyses related to the trial. For the VENT we expect that change in lobar volume will be an indicator of the success of the treatment. Therefore as part of the QC procedures CT-derived lung volumes are compared against total lung capacity obtained from pulmonary function tests (body plethysmography). 3. Results A large number of clinical trial subjects have been analysed within the CAD infrastructure. For example, in the VENT 1264 patients were screened using quantitative CAD. As part of the QC for these subjects, their TLC CT volume was found to have a high correlation (r = 0.89) with pulmonary function tests. Fig. 5 shows the potential of CAD to evaluate lobar volume reduction in emphysema treatment trials. It shows 3D renderings

Fig. 4. Web-based report form showing CAD output: emphysema scores based on density mask and lobe to be targeted for valve placement.

336

M.S. Brown et al. / Computerized Medical Imaging and Graphics 31 (2007) 332–337

Fig. 5. Segmented lobes of the right lung from a thin-section CT exam acquired at total lung capacity: (a) pre-treatment scan and (b) following right upper lobe volume reduction.

of segmented right lung lobes before and after a lobar volume reduction treatment. The images indicate reduction of the right upper lobe volume and expansion of the right lower lobe posttreatment. 4. Discussion This work represents a major advance for CAD in terms of treatment decision support and response assessment. Physicians are now making treatment decisions in a minimally invasive procedure based directly on output from a CAD system. Quantitation of lung disease on CT is active area of research [22,1,3] and in particular for emphysema [10]. A number of systems are under development for automated lobar segmentation on thin-section CT image data [23–25]. However, reliable lobar segmentation remains a challenge because of anatomically incomplete fissures, anatomic variability of the minor fissure and because of irregularities in the fissural plane due to diseases such as emphysema. Moreover, these techniques are only applicable to thin-section imaging. Therefore, to achieve reliable lobar segmentation using thick-section CT images we employed the semi-automated approach described above. To our knowledge, the VENT represents the largest number of cases on which quantitative lobar lung analysis has been performed in a clinical trial. A key to acceptance and wide-spread application of CAD technology is in strong validation studies and an emphasis on

standardization and QC. The density mask measure has been validated against pathology [9], and analyses for QC purposes have been central to VENT as described in Section 3. For CAD to be useable in clinical trials a number of components are required beyond the CAD system itself. In this paper we have described the key components and an architecture in which they are linked through a database management system. We have described an application of the architecture in an emphysema trial where CAD played a new and pivotal role in treatment planning. The architectural requirements are similar for clinical trials involving other diseases and organ systems, and we have also applied this infrastructure in the oncology domain [26,27]. Acknowledgement The authors would like to acknowledge the support of Emphasys Medical, Inc. References [1] Doi K. Overview on research and development of computer-aided diagnostic schemes. Semin Ultrasound CT MR 2004;25(5):404–10. [2] van Ginneken B, ter Haar Romeny BM, Viergever MA. Computer-aided diagnosis in chest radiography: a Survey. IEEE Trans Med Imaging 2001;20:1228–41. [3] Li Q, Li F, Suzuki K, Shiraishi J, Abe H, Engelmann R, et al. Computer-aided diagnosis in thoracic CT. Semin Ultrasound CT MR 2005;26(5):357–63.

M.S. Brown et al. / Computerized Medical Imaging and Graphics 31 (2007) 332–337 [4] Malich A, Fischer DR, B¨ottcher J. CAD for mammography: the technique, results, current role and further developments. Eur Radiol 2006;16(7):1449–60. [5] Summers RM, Jerebko AK, Franaszek M, Malley JD, Johnson CD. Colonic polyps: complementary role of computer-aided detection in CT colonography. Radiology 2002;225:391–9. [6] Wiemker R, Rogalla P, Blaffert T, Sifri D, Hay O, Shah E, et al. Aspects of computer-aided detection (CAD) and volumetry of pulmonary nodules using multislice CT. Br J Radiol 2005;78(1):S46–56. [7] Tran LN, Brown MS, Goldin JG, Yan X, Rogers SR, Pais RC, et al. Comparison of treatment response classifications between unidimensional, bidimensional, and volumetric measurements of metastatic lung lesions on chest CT. Acad Radiol 2004;11(12):1355–60. [8] Reeves AP, Chan AB, Yankelevitz DF, Henschke CI, Kressler B, Kostis WJ. On measuring the change in size of pulmonary nodules. IEEE Trans Med Imaging 2006;25(4):435–50. [9] Muller NL, Staples CA, Miller RR, Abboud RT. Density mask: an objective method to quantitate emphysema using computed tomography. Chest 1988;94:782–7. [10] Coxson HO, Mayo JR, Behzad H, Moore BJ, Verburgt LM, Staples CA, et al. A quantification of the lung surface area in emphysema using computed tomography. Am J Respir Crit Care Med 1999;159(3):851–6. [11] http://www.emphasysmedical.com/wt/emp/clinical trials. [12] Venuta F, de Giacomo T, Rendina E, Ciccone A, Diso D, Perrone A, et al. Bronchoscopic lung-volume reduction with one-way valves in patients with heterogenous emphysema. Ann Thorac Surg 2005;79(2):411–6. [13] Hopkinson NS, Toma TP, Hansell DM, Goldstraw P, Moxham J, Geddes DM, et al. Effect of bronchoscopic lung volume reduction on dynamic hyperinflation and exercise in emphysema. Am J Respir Crit Care Med 2005;171(5):453–60. [14] Yim A, Hwong T, Lee T, Li W, Lam S, Yeung T, et al. Early results of endoscopic lung volume reduction for emphysema. J Thorac Cardiovasc Surg 2004;127(6):1564–73. [15] Snell GI, Holsworth L, Borrill ZL, Thomson KR, Kalff V, Smith JA, et al. The potential for bronchoscopic lung volume reduction using bronchial prostheses: a pilot study. Chest 2003;124(3):1073–80. [16] Wan IY, Geddes TT, Snell DM, Williams G, Venuta T, Yim AP F. Bronchoscopic lung volume reduction for end-stage emphysema: report on the first 98 patients. Chest 2006;129(3):518–26. [17] Zhou Z, Liu BJ, Brown MS, Huang HK, Documet J, Liou D, et al. A Data grid for imaging-based clinical trials. In: Radiological Society of North America scientific assembly and annual meeting program. Oak Brook, Ill: Radiological Society of North America; 2006. p. 795, [abstr]. [18] Brown MS, McNitt-Gray MF, Mankovich NJ, Goldin JG, Hiller J, Wilson LS, et al. Method for segmenting chest CT image data using an anatomical model: preliminary results. IEEE Trans Med Imaging 1997;16(6):828–39. [19] Brown MS, Goldin JG, McNitt-Gray MF, Greaser LE, Sapra A, Li KT, et al. Knowledge-based segmentation of thoracic CT images for assessment of split lung function. Med Phys 2000;27(3):592–8. [20] Ochs R, Abtin F, Brown K, Batra P, Goldin JG, McNitt-Gray MF, et al. A Unified approach for voxel-by-voxel classification of lung anatomy in CT. In: Radiological Society of North America scientific assembly and annual meeting program. Oak Brook, Ill: Radiological Society of North America; 2006. p. 795, [abstr]. [21] Brown MS, McNitt-Gray MF, Wyckoff N, Bui A. Object-oriented region of interest toolkit for workstations, medical imaging 1998: image display. Proc SPIE 1998;3335:627–36. [22] Hoffman EA, Reinhardt JM, Sonka M, Simon BA, Guo J, Saba O, et al. Characterization of the interstitial lung diseases via density-based and texture-based analysis of computed tomography images of lung structure and function. Acad Radiol 2003;10(10):1104–18.

337

[23] Zhang L, Hoffman EA, Reinhardt JM. Lung lobe segmentation by graph search with 3D shape constraints. In: Proceedings of SPIE vol. 4321: medical imaging 2001: physiology and function from multidimensional images. 2001. p. 204–15. [24] Kuhnigk JM, Hahn HK, Hindennach M, Dicken V, Krass S, Peitgen HO. Lung lobe segmentation by anatomy-guided 3D watershed transform. In: Proceedings of SPIE vol. 5032 medical imaging 2003: image processing. 2003. p. 1482–90. [25] Zhou X, Kobayashi S, Hayashi T, Murata N, Hara T, Fujita H, et al. Lung structure recognition: a further study of thoracic organ recognitions based on CT images. In: Proceedings of CARS 2003. International congress series 1256. 2003. p. 1025–30. [26] Petkovska I, Shah SK, McNitt-Gray MF, Goldin JG, Brown MS, Kim HJ, et al. Pulmonary nodule characterization: a comparison of conventional with quantitative and visual semi-quantitative analyses using contrast enhancement maps. Eur J Radiol 2006;59(2):244–52. [27] Brown MS, McNitt-Gray MF, Goldin JG, et al. Patient-specific models for lung nodule detection and surveillance in CT images. IEEE Trans Med Imaging 2001;20(12):1242–50. Matthew Brown received his PhD in computer science from the University of New South Wales in 1997. His PhD research was in knowledge-based interpretation of medical images. In 1998 he joined the faculty in the UCLA Department of Radiological Sciences where he is currently an associate professor and continues to pursue his research in computer vision and computer-aided diagnosis from medical images. Michael McNitt-Gray, PhD is an associate professor in the Department of Radiological Sciences at the David Geffen School of Medicine at UCLA, and is director of the Biomedical Physics Graduate program there. Dr. McNittGray received the BSEE from Washington University in St. Louis, MSEE from Carnegie-Mellon University and the PhD from UCLA in biomedical physics. Dr. McNitt-Gray’s research interests involve investigations into X-ray computed tomography (CT) imaging with the goal of maximizing the information that can be extracted from the resulting image data. These activities include research into the physics of CT image acquisition including estimating radiation dose and assessing image quality and image processing techniques to analyze and extract information from the CT image data, including computer aided detection and diagnosis (CAD). Richard Pais has been actively developing code in Java for the past 7 years using open source technologies. His primary interest is in software engineering and architecture, database design, modeling and optimization techniques and XML technologies. Dr. Denise Aberle is a professor in Thoracic Imaging/Radiological Sciences in the David Geffen School of Medicine at UCLA and in the Department of Bioengineering in the Henry Samueli School of Engineering and Applied Sciences. Board certified in internal medicine and diagnostic radiology, her interests are in (a) lung cancer and oncologic imaging technologies for both early detection and response assessment; (b) quantitative image analysis, and (c) oncology informatics. Jonathan Goldin is a professor of radiology at the David Geffen School of Medicine at UCLA. He is clinical vice chair of the Department of Radiology, chief of Radiology at the Santa Monica-UCLA Medical Center and Division Chief of the Santa Monica-UCLA Multi-Specialty Radiology Department in the David Geffen School of Medicine. He is also a director of both the Thoracic Imaging Research Laboratory and the Thoracic Imaging Fellowship Program. Dr. Goldin’s clinical and research interests include use of imaging in the early detetction of disease, including lung cancer and coronary calcification screening as well as the development of computer-aided systems for the detection, characterization and quantization of diffuse lung disease and cardiovascular diseases.