A computer-aided temporal and dynamic subtraction technique of the liver for detection of small hepatocellular carcinomas on abdominal CT images

It is often difficult for radiologists to identify small hepatocellular carcinomas (HCCs) due to insufficient contrast enhancement. Therefore, we have developed a new computer-aided temporal and dynamic subtraction technique to enhance small HCCs, after automatically selecting images set at the same anatomical position from the present (non-enhanced and arterial-phase CT images) and previous images. The present study was performed with CT images from 14 subjects. First, we used template-matching based on similarities in liver shape between the present (non-enhanced and arterial-phase CT images) and previous arterial-phase CT images at the same position. Temporal subtraction images were then obtained by subtraction of the previous image from the present image taken at the same position of the liver. Dynamic subtraction images were also obtained by subtraction of non-enhanced CT images from arterial-phase CT images taken at the same position of the liver. Twenty-one of 22 nodules (95.5%) with contrast enhancement were visualized in temporal and dynamic subtraction images. Compared with present arterial-phase CT images, increases of 150% and 140% in nodule-to-liver contrast were observed on dynamic and temporal subtraction images, respectively. These subtraction images may be useful as reference images in the detection of small moderately differentiated HCCs.


Introduction
The number of patients with hepatocellular carcinomas (HCCs) is increasing worldwide. Therefore, patients in high-risk groups consisting of those with liver cirrhosis and hepatitis B and C undergo routine screening (Farinati andGianni 2001, Ward andRobinson 2002). Although ultrasonography (US) is widely used in screening for HCCs, it has limitations in the detection of small tumours (Kim et al 2001). Therefore, abdominal CT examination is also performed (Ward and Robinson 2002). However, hypervascular tumours are often located adjacent to vascular structures and are often minute. It is especially difficult for radiologists to distinguish true HCCs from small nodular arterioportal shunts (AP shunts) in patients with liver cirrhosis. Therefore, the sensitivity of CT examination for detection of HCCs is low, ranging from only 59-68% (Peterson et al 2000, Baron and Peterson 2001, Blachar et al 2002. Hypervascular moderately differentiated HCCs usually exhibit early enhancement on arterial-phase CT images and ring enhancement on delayed-phase CT images. However, it is necessary to differentiate HCCs from small AP shunts, because the latter often show nodular enhancement on arterial-phase CT images. Therefore, the development of methods for computer-aided diagnosis (CAD) is required to avoid overlooking nodular early enhancement and to increase the reliability of diagnostic imaging of HCCs.
A number of groups have reported dynamic subtraction techniques for US, CT and MRI of the liver region (Soyer et al 1999, Ding et al 2001, Spielmann et al 2002, Yu and Rofsky 2003. However, use of these techniques in routine clinical practice requires the selection of US, CT and MR images at the same anatomical position, increasing both the time required and cost as radiologists must visually select images at the same positions of the liver. In a preliminary study, we developed an automated image registration technique to allow the selection of image sets corresponding to the same anatomical positions from non-enhanced and arterial-phase CT images (Okumura et al 2005). Using this automated image registration technique, we have developed a new dynamic subtraction technique to detect early enhancement, including moderately differentiated HCCs and AP shunts.
A number of groups have also reported temporal subtraction techniques for chest CT images and posteroanterior (PA) chest images (Kano et al 1993, Ishida et al 1998, Abe et al 2004. These investigators reported that temporal subtraction techniques enhanced many subtle changes in chest CT images and PA chest images. However, to our knowledge, there have been no previous reports regarding the detection of HCCs by temporal subtraction. Therefore, we have developed a new temporal subtraction technique to detect increases in size of nodular early enhancement and new enhancement, after automatically selecting images set at the same anatomical positions from present and previous CT images.

Materials
The study population consisted of fourteen patients with HCCs (ten men and four women; mean age, 65.1 years). Pairs of images (previous and present images) were obtained for each patient. Each patient underwent non-enhanced and arterial-phase CT imaging of the upper abdomen using a LightSpeed Utra16 scanner (GE Yokogawa Medical Systems, Tokyo, Japan; 120 kVp and 200 mA). The matrix size of the images was 512 × 512 pixels, and the slice thickness and number of images in 14 examinations were 1.25-3.75 mm and 65-200 images, respectively. The previous and present images were obtained from October 2002 to June 2004, and the interval between images was 174 ± 80.5 days (range, 84-377 days). Based on prospective interpretation of the present images, radiologists suspected 18 HCCs and one AP shunt, and missed five tumours. Based on the previous images, radiologists suspected six HCCs and six AP shunts. Twenty-four nodules, including the five missed tumours, were diagnosed as HCCs on angiography, CT hepatic arteriography (CTHA) and CT arterial portography (CTAP) performed after the present CT examination.

Development of the computer algorithm
The application described here was developed using Borland C++ Builder 6 (Borland, Scotts Valley, CA, USA). The CT images were transferred off-line from a CT scanner to a personal computer (PC) with a 2.0 GHz CPU.

Detection of the liver region
The right hepatic lobe occupies almost the entire right half of the abdomen on CT images, and the shape of the liver region changes gradually in the cranial to caudal direction. However, it is often difficult to extract all liver regions using methods based only on detecting the maximum area, because the left and right lobes may be visualized separately, and the spleen and other organs may have areas larger than the liver in some images.
The image centre of gravity in the liver region does not show marked variation in the neighbouring images. We attempted to detect the liver region in all images, taking into consideration the centre of gravity in the top and bottom images.
The present non-enhanced and arterial-phase CT images and the previous arterial-phase CT images were smoothed. A rectangular region of interest (ROI) was set manually at the centre of the liver region on the reference image, in the middle of the craniocaudal portion of the liver. We obtained the threshold value of the liver in this ROI by histogram analysis and binarized all images using this threshold value. These binary images were obtained where a pixel was assigned a value of 0 or 1 if the density was below or above the threshold value, respectively. We employed opening processing (erosion and dilation) to eliminate the extrahepatic structures, such as the abdominal wall and stomach, and the centre of gravity of the labelled regions was calculated after labelling. The labelled regions of the adjacent images, including the centre of gravity of the image, were then detected as the liver region (Okumura et al 2005).

Image matching of the same anatomical positions
Image matching has been carried out to identity the posteroanterior and lateral views of chest radiographs (Arimura et al 2002) and for patient recognition of previous and current chest radiographs (Morishita et al 2001(Morishita et al , 2004 using intensity-based methods, which make use of residual and correlation coefficients, etc. The residual, R, which indicates the extent of similarity between template A(i, j ) and training images B(i, j ), was determined as follows: (1) The matrix size for the template and training images was M × N . By comparison of the residual between template and training images, the template images can be considered as training images that provide lesser residuals. In the present study, we used the residual for image matching of the same anatomical positions, because we produced binarized arterialphase CT images as template images and binarized non-enhanced CT images as training images, and it was useful to show image subtraction with the sum of intensity differences.

Dynamic subtraction technique
First, we produced binarized present arterial-phase CT images as template images, and binarized non-enhanced CT images as training images. To register the present non-enhanced and arterial-phase CT images obtained in the same period, we used template matching (automated image matching) based on similarities in liver shape between the present nonenhanced and arterial-phase CT images at the same position. Automated image-matching was then applied to determine the global shift value ( x, y) between the two images showing the best match, which corresponds to the shift in the x, y coordinates of the template images relative to the shift in the x , y coordinates of training images (Ishida et al 1998) using equation (2): (2) After performing the global shift ( x, y), dynamic subtraction images were obtained by subtraction of non-enhanced CT images from arterial-phase CT images taken at the same position of the liver.

Temporal subtraction technique
First, we produced binarized present arterial-phase CT images as template images, and the binarized previous arterial-phase CT images as training images. Then, we registered the previous and present arterial-phase CT images at the same position in the same manner as described for dynamic subtraction. After performing the global shift ( x, y), temporal subtraction images were obtained by subtraction of the previous images from the present images taken at the same position of the liver.

Subjective evaluation of the quality of dynamic and temporal subtraction images
All cases were assessed by three experienced radiologists after consultation with no knowledge of the presence of HCCs in all cases. Four sets of CT images were evaluated for the presence of HCCs, and were classified as: (a) presence of HCC, (b) suspected HCC, and (c) no HCC. Set A (only present non-enhanced CT images) was assessed first, followed by set B (present non-enhanced and arterial-phase CT images), set C (non-enhanced, arterial-phase, and previous arterial-phase CT images), and set D (set C plus dynamic and temporal subtraction images), and this cycle was then repeated. In addition, the image quality of dynamic and temporal subtraction images indicating confirmed or suspected HCCs in set C was assessed by three radiologists using a three-point score: 3 (good), whole nodule was visualized with minor misregistration; 2 (adequate), nodule was visualized with some misregistration; and 1 (poor), nodule was not detected with marked misregistration error. Finally, the usefulness of both subtraction images in detecting HCC was classified as follows: +1, very helpful; 0, no additional information; and −1, not helpful.

Objective evaluation of the quality of dynamic and temporal subtraction images
To allow objective evaluation of the quality of dynamic subtraction images using the present CT images and temporal subtraction images, we calculated the nodule-to-liver contrast (Contrast) and signal-to-noise ratio (SNR) of the present arterial-phase CT images and dynamic subtraction images using the present CT images and temporal subtraction images (Biswas et al 2005) according to equations (3) and (4) We selected regions of interest (ROIs) for the liver and the lesion. Five ROIs of about 10 × 10 pixels were deposited manually in the liver region. We measured the average value of five ROIs, which was defined as CT liver of the liver region. The boundary of the nodule was traced manually to measure the ROI for the nodule region. We measured CT nodule in the nodule region in the same manner as described for measurement of CT number in the liver region. We measured the standard deviation (SD air ) of the background to estimate image noise. Respective data for Contrast and SNR between the present arterial-phase CT images and temporal subtraction images, and between the present arterial-phase CT images and dynamic subtraction images using the present CT images with good rating scores were analysed using Wilcoxon's two-tailed paired t-test, and P 0.05 was considered statistically significant.

Results
Three experienced radiologists detected 21 of 24 HCCs (87.5%), suspected a HCC in one case, and missed two HCCs in sets A-C. Twenty-two of 24 HCCs (91.6%) were detected, a HCC was suspected in one case, and one HCC was not detected in set D. There were no incidences of false positives. The temporal and dynamic subtraction images enhanced the detection of HCCs. With regard to image quality of dynamic and temporal subtraction images, 21 HCCs were classified as good and one HCC as poor. With regard to the image quality of dynamic subtraction images alone, 20 HCCs showing nodular early enhancement were classified as good, while in the image quality of temporal subtraction image alone, 16 HCCs showing ring or nodular enhancement were classified as good.
On subtraction images, none of seven HCCs with nodular enhancement on the present images showed abnormal enhancement on the previous images. All seven HCCs showing nodular enhancement on the dynamic subtraction images were classified as good, while five of seven HCCs showing nodular enhancement on the temporal subtraction images were classified as good (figure 1). Fourteen HCCs suspected to be nodular lesions on the previous images showed increases in size on the present images. Twelve of 14 HCCs showing nodular enhancement on dynamic subtraction images were classified as good, while 10 of 14 HCCs showing ring enhancement on temporal subtraction images were classified as good (figure 2). One HCC was suspected based on previous and present images; enhanced lesions were seen on both temporal and dynamic subtraction images in this HCC and the results were classified as good (figure 3). With regard to the usefulness of subtraction images, 3 of 21 HCCs classified as good were classified as very helpful. The subtraction images increased the conspicuousness of detection of HCC ( figure 4).
However, one HCC detected in sets A-C was not visualized on both subtraction images (figure 5), and was classified as poor.
The averages and SD of nodule-to-liver contrast and nodule-to-liver contrast ratio of the present arterial-phase CT images and temporal subtraction images, and dynamic subtraction images using the present CT images are shown in table 1. These temporal and dynamic subtraction images using the present CT images that were rated as good showed increases in nodule-to-liver contrast of 140% and 150%, respectively, compared to the present arterialphase CT images (P < 0.01). In addition, the averages and SD of SNR and ratio of SNR of the present arterial-phase CT images and temporal subtraction images, and dynamic subtraction images using the present CT images of the liver and nodule region are also shown in table 1. For good rating score of dynamic subtraction images using the present CT images and temporal subtraction images compared with the present arterial-phase CT images, the former showed deterioration in SNR, respectively (P < 0.01). In cases in which both subtraction images were mostly registered with only minor misregistrations, nodule-to-liver contrast of both subtraction images was higher than those of the present arterial-phase CT images, and SNR of both subtraction images was lower than those of the present arterial-phase CT images in the liver and nodule region.

Discussion
It is often difficult for radiologists to detect HCCs adjacent to vascular structures and to distinguish them from other small hypervascular tumours or small nodular AP shunts (Peterson et al 2000, Baron and Peterson 2001, Blachar et al 2002  or new enhancement on temporal subtraction images and were classified as good. In addition, dynamic subtraction images were useful in assisting in the detection of early enhancement, such as moderately differentiated HCCs and AP shunts. Temporal subtraction images showed an increase of ring enhancement in the lesions in the previous arterial-phase CT images that indicated the presence of HCCs and nodular enhancement showed the appearance of new lesions. In the present study, 21 of 24 HCCs were detected on the present non-enhanced, arterial-phase and the previous arterial-phase CT images, while 22 HCCs were detected using the additional information from subtraction images. In addition, the dynamic and temporal subtraction images increased the conspicuousness of HCCs. On objective evaluation of the quality of dynamic and temporal subtraction images, as the levels of background noise of both subtraction images were higher than those of the present arterial-phase CT images, SNRs of both subtraction images were lower than those of the present arterial-phase CT images. However, both subtraction images had uniform zero pixel values except for regions with interval changes, etc. Therefore, nodule-to-liver contrast was higher in both subtraction images than in present arterial-phase CT images. In the present study, there was a difference between HCCs classified as good in the dynamic and temporal subtraction images. The difference was mainly due to misregistration of the hepatic shape, which was related only to breath-holding on the dynamic subtraction images, while it was related to breath-holding and hepatic deformity during the clinical course on temporal subtraction images. There have been a number of studies of computer-aided detection of HCCs (Bellon et al 1997, Voirin et al 2002, Gletsos et al 2003, Bilello et al 2004. Gletsos et al (2003) reported a method for detecting and classifying HCCs using texture analysis and an artificial neural network (ANN) on non-enhanced CT images. Bilello et al (2004) also reported a method for detecting HCCs on delayed-phase CT images. In their study, to achieve 90% sensitivity, the total false-positive rate was increased to about 2.2 per section. In the present study, after assessing set C (non-enhanced, arterial-phase and previous arterial-phase CT images), observers then assessed both subtraction images. Therefore, there were no false-positives due to the enhanced images. Screening for HCCs by abdominal CT is widely performed in patients with liver cirrhosis and hepatitis B and C. In general, there are no remarkable changes in liver shape during such screening intervals, thus reducing the frequency of misregistration artefacts. It is convenient to apply temporal subtraction for detection of HCCs under such conditions. Therefore, this technique using not only dynamic subtraction but also temporal subtraction may be beneficial to radiologists for CT detection of HCCs. However, one nodule was not detected on both dynamic subtraction images using the present CT images and temporal subtraction images. As shown in figure 5, the radiologist could not locate this tumour precisely on either image. In this case, different axial levels of the liver were selected in the automated image-matching technique. Therefore, misregistration artefacts appeared at the periphery of the liver. In addition, no definite staining characteristic of HCCs was observed, probably because of the relatively hypovascular nature of the lesion. For good rating scores, some misregistration artefacts appeared at the periphery of the liver in dynamic subtraction images using the present CT images and temporal subtraction images (figures 1-4). With regard to the manual global shift value, we obtained dynamic subtraction images using the present CT images and temporal subtraction images in these cases, after manual selection of image sets at the same anatomical positions from the present (non-enhanced and contrast-enhanced images) and previous images. Less misregistration artefacts appeared at the periphery of the liver in these images than in the automatic method.
The present study did not include cases with deformity of the liver due to treatment, such as transcatheter arterial embolisation (TAE), percutaneous ethanol injection therapy (PEIT) or radiofrequency ablation (RFA). In addition, changes in hepatic shape may be induced by local atrophy of the liver between the present and previous images, ascites, or breath-holding. Therefore, it may be difficult to select images from the same anatomical positions in such cases. It is necessary to improve the precision of this automated image-matching technique for use in such cases. Therefore, in future studies we intend to compare the volume data of the liver on each of the previous and present images. Methods involving not only image shifting and rotation techniques but also nonlinear geometric warping techniques for previous images (Kano et al 1993, Ishida et al 1999 may improve the method developed here because of the reduction of misregistration artefacts, especially on temporal subtraction images. The results were obtained in a relatively small number of cases (14 cases) and with a small number of observers (three observers). Therefore, to evaluate the clinical efficacy of this technique, a prospective study (ROC type analysis) with larger numbers of patients is required.

Conclusions
We developed a computerised method for detecting moderately differentiated HCCs using dynamic and temporal subtraction images based on an automated image-matching technique. This method may be useful for use of these subtraction images as reference images in detection of small moderately differentiated HCCs by CT.