First published online July 19, 2006
British Journal of Radiology (2007) 80, 85-89
© 2007 British Institute of Radiology
doi: 10.1259/bjr/29418923
Evaluation of an automated system for temporal subtraction of thin-section thoracic CT
H Takao, MD1,2,
I Doi, MD1 and
M Tateno, MD2
1 Department of Radiology, Showa General Hospital, Tokyo and, 2 Department of Radiology, Social Insurance Central General Hospital, Tokyo, Japan
Correspondence: Hidemasa Takao, Department of Radiology, Showa General Hospital, 2-450 Tenjincho, Kodaira, Tokyo 187-0004, Japan. E-mail: takaoh-tky{at}umin.ac.jp
 |
Abstract
|
|---|
The purpose of this study was to objectively evaluate the registration accuracy of an automated temporal subtraction system of thin-section thoracic CT. The automated subtraction system was applied to data from 20 subjects with lung diseases. The registration accuracy was evaluated based on the concept of target registration error using 19 landmarks chosen at the bifurcations of segmental bronchi. The landmarks were identified, and the displacements of the landmarks were measured. The average displacements of the landmarks in x, y and z directions were 0.56 pixels±0.64 (SD), 0.56 pixels±0.62 (SD, standard deviation) and 0.28 sections±0.40 (SD), respectively. Differences in average displacements between the landmarks were not statistically significant. Our computer system demonstrated promising results. The average displacements of the landmarks were less than the dimensions of a voxel. Further study is necessary to determine whether such a computer system can improve the diagnostic performance of radiologists.
|
Introduction
|
|---|
With recent advances in CT technology, spatial resolution in the longitudinal plane has dramatically improved. Multidetector-row CT (MDCT) allows the acquisition of a volumetric data set within a single patient breath-hold. With the latest 64-detector row CT scanners, the whole thorax can be imaged in a few seconds. The availability of thin sections enables the detection and characterization of small abnormalities. However, each MDCT study results in a large data set, typically 130–170 images of 2 mm section thickness, and requires radiologists to spend a considerable amount of time interpreting the images. The detection of interval changes on thoracic CT is an important task for radiologists. This task has become more difficult and more time-consuming as we migrate to increasingly higher levels of longitudinal spatial resolution and larger data sets with MDCT. Therefore, we have developed a computer system that automatically produces temporal subtraction images on thin-section thoracic CT based on a three-dimensional non-linear geometric warping technique [1]. The purpose of this study was to objectively evaluate the registration accuracy of the automated temporal subtraction system of thin-section thoracic CT.
|
Methods and materials
|
|---|
Patient selection
The approval of our Institutional Review Board was not required for this retrospective study. Informed consent was not required. 20 patients (19 men and 1 woman; age range 50–88 years; mean age 64.7 years) with lung diseases were selected from the patients who underwent thin-section thoracic CT at our institute for clinical indications from October to November 2004 [1]. A patient was excluded if the reports of CT studies included the terms "consolidation" or "collapse", if the patient had a pleural effusion, or if the patient had not undergone previous thin-section CT at our institution. Each patient's clinical course and diagnosis was confirmed by means of chart review. Patient diagnoses were lung carcinoma (n = 7), pulmonary metastases (n = 2), tuberculosis (n = 1), sarcoidosis (n = 1) aspergilloma (n = 1), rounded atelectasis (n = 1) and old scarring (n = 7). The mean interval between CT examinations was 5.5 months (range 1–15 months).
Imaging protocol
All examinations were performed with a MDCT scanner (Asteion/Toshiba, Tokyo, Japan) with the following parameters: collimation 4 mmx2.0 mm; pitch 1.375 (corresponding to manufacturer's pitch of 5.5) [2]; rotation time 0.75 s; 120 kVp; and 200 mAs. The thorax was scanned in a caudocranial direction. The CT data were reconstructed into 2.0 mm transverse sections at 2.0 mm intervals, and a high-frequency reconstruction algorithm and a matrix of 512x512 pixels were used. The displayed field of view ranged between 30–36 cm.
Temporal subtraction
The temporal subtraction technique on thin-section thoracic CT has been described in detail elsewhere [1]. The temporal subtraction system that we developed was implemented on a personal computer (Dell Precision WorkStation 360; Dell, Round Rock, TX) with a 2.4 GHz Pentium 4 processor, 1 GB of RAM and a Windows XP Professional operating system.
First, the computer system excluded the extrapulmonary compartment from each CT image. Initial image matching was performed to roughly determine shift values between current and previous images with low-resolution images. Subsequently, multiple square template regions of interest (ROIs) were automatically placed on the current images. Local matching of template ROIs with the corresponding search area ROIs in the previous images was performed using a cross-correlation method [3, 4]. Shift values 
x, y and z were determined for all pairs of ROI. Three-dimensional polynomial curves were then fitted to the shift values by using tenth-order polynomials based on a least-square method. Non-linear geometric warping of the x,y,z coordinates of the previous images was performed with the fitted shift values. Finally, the warped (three-dimensionally registered) previous images were subtracted from the current images (
Figures 1 and 2).
View larger version (73K):
[in this window]
[in a new window]
|
Figure 1. 51-year-old man with sarcoidosis. (a) Current CT image shows a focal area of mass-like consolidation with air-filled cavities in the right lung. Multiple nodules and nodular thickening in the peribronchovascular bundles are also seen in both lungs. (b) Temporal subtraction image shows increase in size of the air-filled cavities (arrows). Black areas indicate increase in size of the cavities. (c) Warped previous CT image (the extrapulmonary compartment is excluded). Most structures of both lungs are well matched with those of the current CT image. (d) Original previous CT image. Although the lesions in the right lung are matched with those of the current image, most structures of the left lung are not matched with those of the current CT image. Registration is performed three-dimensionally (non-rigidly). (c) The warped previous CT image was produced from the previous CT images, not only from (d) the single previous CT image.
|
|
View larger version (74K):
[in this window]
[in a new window]
|
Figure 2. 63-year-old man with lung cancer. (a) Current CT image shows several pulmonary nodules in the left upper lung. (b) Temporal subtraction image shows new appearance of mucoid impaction in a bronchus (arrow) of the left upper lung, which is depicted as a bright area. (c) Warped previous CT image (the extrapulmonary compartment is excluded). (d) Original previous CT image.
|
|
Processing computation time
The computer system was completely automated and no user input was required. It took approximately 5 min (range 3.4–6.3 min; mean 5.1 min) to produce subtraction images from the original images (for the whole data set).
Evaluation of registration
Image evaluation was performed on a personal computer (Dell Precision WorkStation 360) with a Windows XP Professional operating system and a diagnostic-quality picture archiving and communication systems (PACS) monitor (RadiForce G20, Nanao, Ishikawa, Japan). The monitor had a luminance of 700 cd m–2, an aperture grille pitch of 0.255 mm, a resolution of 1600x1200 and a refresh rate of 60 Hz.
We evaluated the registration accuracy based on the concept of target registration error using landmarks [4–6]. 19 landmarks were chosen at the bifurcations of segmental bronchi (Table 1) [7]. A radiologist (HT) with 6 years experience identified the landmarks and measured the displacements of the landmarks. The radiologist identified the landmark on the current images, obtained an (x,y,z) position for it, then identified the same landmark in the warped previous images and obtained an (x,y,z) position for it in the same coordinate system as for the current images. The displacements were then calculated. All measurements were performed twice and results were averaged to minimize measurement-related errors.
Statistical analysis
Statistical analyses were performed with a commercially available computer program (SPSS for Windows, version 11.0.1; SPSS Inc., Chicago, IL). All data are presented as means±standard deviations (SDs). One-way analysis of variance (ANOVA) was used to test differences in average displacements between the landmarks. A p-value of less than 0.05 was considered to indicate a statistically significant difference.
|
Results
|
|---|
In one patient, one landmark (right B1) could not be identified because of tumour invasion. In another patient, four landmarks (left B6–B10) could not be identified because of previous surgery. The average displacements of the landmarks in x, y and z directions were 0.56±0.64 pixels, 0.56±0.62 pixels, and 0.28±0.40 sections, respectively. Differences in average displacements between the landmarks were not statistically significant (x direction, p = 0.45; y direction, p = 0.72; z direction, p = 0.70). Further details are given in Table 2.
|
Discussion
|
|---|
In this study, we objectively evaluated the registration accuracy of an automated temporal subtraction system of thin-section thoracic CT based on the concept of target registration error using landmarks chosen at the bifurcations of segmental bronchi. Although our computer analysis is still preliminary, the results of this study are promising. The average displacements of the landmarks were less than the dimensions of a voxel.
There is a growing interest in the development of computer-aided detection (CAD) technology for the early detection of lung cancer with chest radiography and CT [8]. Nodule detection is one of the more challenging, albeit common, tasks in medical imaging. Although double readings reduce the number of missed nodules, routine double reading is unrealistic in clinical practice. Computerized nodule detection schemes have been shown to substantially increase diagnostic accuracy in medical imaging [8]. Although the assessment of nodule growth is important, most CAD systems are not able to assess interval changes between separate studies. The radiology literature reports many registration techniques for the brain and other organ systems. However, there are only a few reports regarding registration of thoracic CT images between different studies [9–12]. The computer system of Ko et al [9] automatically identifies nodules on thoracic CT, quantifies their diameter and assesses for change in size at follow-up. Brown et al [10] developed a lung nodule identification system capable of automated detection and automated follow-up of small nodules on thin-section CT. Dougherty et al [11] applied an optical flow method to register thoracic CT. They used manually selected subregions of thick-section (5–7 mm) thoracic CT images. Abe et al [12] applied the temporal subtraction scheme developed for chest radiographs to thick-section (10 mm) thoracic CT images. The selection of the corresponding section in two sets of CT images and image shift correction between current and previous images were performed manually.
There were several limitations in this study. First, we excluded patients in whom the reports of CT studies included the terms "consolidation" or "collapse", or who had pleural effusions. The aim of our computer system, however, was to assist radiologists in identifying interval changes that are difficult or time-consuming to detect. Thus, we believe that the data used were adequate to evaluate the computer system for this purpose. Second, in this study we evaluated the registration accuracy of the computer system, but did not evaluate its effect on diagnostic performance of radiologists. Further study will be necessary to determine clinical utility of the computer system. Finally, the 5 min computation time for producing temporal subtraction images is relatively long for routine applications. This time, however, will be shortened with the anticipated progress of computer technology.
In conclusion, our computer system demonstrated promising results. The average displacements of the landmarks were less than the dimensions of a voxel. Further study is necessary to determine whether such a computer system can improve the diagnostic performance of radiologists.
Received for publication March 21, 2006.
Revision received May 24, 2006.
Accepted for publication June 20, 2006.
|
References
|
|---|
- Takao H, Doi I, Watanabe T, Tateno M. Temporal subtraction of thin-section thoracic CT based on a three-dimensional nonlinear geometric warping technique. J Comput Assist Tomogr 2006;30:283–6.[CrossRef][Medline]
- Silverman PM, Kalender WA, Hazle JD. Common terminology for single and multislice helical CT. AJR Am J Roentgenol 2001;176:1135–6.[Free Full Text]
- Kano A, Doi K, MacMahon H, Hassell DD, Giger ML. Digital image subtraction of temporally sequential chest images for detection of interval change. Med Phys 1994;21:453–61.[CrossRef][Medline]
- Hill DL, Batchelor PG, Holden M, Hawkes DJ. Medical image registration. Phys Med Biol 2001;46:R1–R45.[CrossRef][Medline]
- Gee J, Sundaram T, Hasegawa I, Uematsu H, Hatabu H. Characterization of regional pulmonary mechanics from serial magnetic resonance imaging data. Acad Radiol 2003;10:1147–52.[CrossRef][Medline]
- Takao M, Sugano N, Nishii T, Tanaka H, Masumoto J, Miki H, et al. Application of three-dimensional magnetic resonance image registration for monitoring hip joint diseases. Magn Reson Imaging 2005;23:665–70.[CrossRef][Medline]
- Johnson D, Shah P, Collins P, Wigley C, editors. Pleura, lungs, trachea and bronchi. In: Standring S, Ellis H, Healy JC, Johnson D, Williams A, Collins P, et al, editors. Gray's anatomy, 39th edition. London, UK: Elsevier, 2005: 1063–79
- Ko JP, Naidich DP. Lung nodule detection and characterization with multislice CT. Radiol Clin North Am 2003;41:575–97.[CrossRef][Medline]
- Ko JP, Betke M. Chest CT: automated nodule detection and assessment of change over time - preliminary experience. Radiology 2001;218:267–73.[Abstract/Free Full Text]
- Brown MS, McNitt-Gray MF, Goldin JG, Suh RD, Sayre JW, Aberle DR. Patient-specific models for lung nodule detection and surveillance in CT images. IEEE Trans Med Imaging 2001;20:1242–50.[CrossRef][Medline]
- Dougherty L, Asmuth JC, Gefter WB. Alignment of CT lung volumes with an optical flow method. Acad Radiol 2003;10:249–54.[CrossRef][Medline]
- Abe H, Ishida T, Shiraishi J, Li F, Katsuragawa S, Sone S, et al. Effect of temporal subtraction images on radiologists' detection of lung cancer on CT: results of the observer performance study with use of film computed tomography images. Acad Radiol 2004;11:1337–43.[CrossRef][Medline]
Top
Abstract
Introduction
