Statistical and Economical Efficiency in Assessment of Liver Regeneration Using Defined Sample Size and Selection in Combination With a Fully Automated Image Analysis System

Abstract

Quantification of liver regeneration is frequently based on determining the 5-bromo-2-deoxyuridine labeling index (BrdU-LI). The quantitative result is influenced by preanalytical, analytical, and postanalytical variables such as the region of interest (ROI). We aimed to present our newly developed and validated automatic computer-based image analysis system (AnalySIS-Macro), and to standardize the selection and sample size of ROIs. Images from BrdU-labeled and immunohistochemically stained liver sections were analyzed conventionally and with the newly developed AnalySIS-Macro and used for validation of the system. Automatic quantification correlated well with the manual counting result (r=0.9976). Validation of our AnalySIS-Macro revealed its high sensitivity (>90%) and specificity. The BrdU-LI ranged from 11% to 57% within the same liver (32.96 ± 11.94%), reflecting the highly variable spatial distribution of hepatocyte proliferation. At least 2000 hepatocytes (10 images at 200× magnification) per lobe were required as sample size for achieving a representative BrdU-LI. Furthermore, the number of pericentral areas should be equal to that of periportal areas. The combination of our AnalySIS-Macro with rules for the selection and size of ROIs represents an accurate, sensitive, specific, and efficient diagnostic tool for the determination of the BrdU-LI and the spatial distribution of proliferating hepatocytes. (J Histochem Cytochem 57:1075–1085, 2009)

5-Bromo-2-deoxyuridine (BrdU) is an analog of thymidine. It can be incorporated into the newly synthesized DNA instead of thymidine during the S phase of the cell cycle. Since deFazio et al. (1987) reported a method for detecting proliferating cells in situ using anti-BrdU immunohistochemical staining, this procedure has been widely used for quantifying cellular proliferation (Dolbeare 1996).

Traditionally, sections underwent reaction with anti-BrdU antibody and the BrdU-labeled and non-labeled cells were counted by microscopists prior to calculating the BrdU labeling index (BrdU-LI) (BrdU-LI = ratio of BrdU-positive cells to total cells in percent). This procedure is time-consuming and tedious and leads to a high work load when a complete animal experiment needs to be evaluated.

The BrdU-LI is one example of tissue-based quantification of cellular events. These results are influenced by a number of variable factors existing on the preanalytical, analytical, and postanalytical level.

Preanalytical variables include variations in the sampling procedures, sample selection, fixation, paraffinization, and section cutting. Analytical variables include differences in the staining procedures. Standardization of the sampling and staining procedures in detail for each organ helps to reduce the preanalytical and analytical variables, and thereby reduce the sampling error and the variations in the staining intensity (Ruehl-Fehlert et al. 2003; Nolte et al. 2005).

Postanalytical variables include the experience and bias of the observer in identifying the target events as well as the selection and size of regions of interest (ROIs) to be evaluated. The sampling bias generated by the selection and sample size of ROIs is crucial in the case of inhomogeneous spatial distribution of the targeted cellular event, such as hepatocyte proliferation in the liver (Gebhardt and Jonitza 1991; Chen et al. 1995). Great effort was paid to improve both statistical efficiency (the term “statistical efficiency” refers to the precision of an estimator) and economical efficiency (the term “economical efficiency” refers to the workload for achieving a given precision of the estimates) in the assessment of liver regeneration. Computer-assisted quantification of a large and representative ROI is the ultimate way to minimize observer-related bias and to improve the economical efficiency (Soames et al. 1994). The standardization of the selection and sample size of an ROI is one way to minimize the sampling bias based on two-dimensional measurement (Bahnemann and Mellert 1997). Design-based stereology is a novel way to minimize the sampling bias based on three-dimensional measurement.

In the present study, we aimed to present our newly developed and validated computer-based fully automatic AnalySIS-Macro, which eliminates the observer-related bias and greatly improves the economical efficiency; and to standardize the selection and size of ROIs to be analyzed for assessing hepatocyte proliferation in liver regeneration, based on the use of the computer-based fully automatic AnalySIS-Macro.

Materials and Methods

Animals

Male inbred Lewis rats weighing 250 g to 350 g were purchased from the animal facility of University Hospital Essen. Animals were housed under standard animal care conditions and fed with rat chow ad libitum. All procedures were carried out in accordance with German animal welfare legislation.

Surgical Treatment

Seventy percent partial hepatectomy (PH) was performed under inhalation anesthesia with 1.5–3% isoflurane (Sigma; St. Louis, MO). The left lateral lobe (∼30%) and the whole median lobe (∼40%) were resected, leading to an estimated 70% reduction of liver mass (Madrahimov et al. 2006). Postoperative analgesia was achieved by subcutaneous injection of buprenorphine (0.01 mg/kg) (TemgesicTM; Essex Pharma, Munich, Germany). Animals were sacrificed 24 hr postoperatively. One animal (MAR001) was sacrificed 24 hr postoperatively and underwent multiple sampling from each remnant liver lobe. Sampling was limited to the superior caudate lobes in four additional rats subjected to 70% PH and sacrificed at 24 hr after resection (animals AEE015, -016, -017, and -018). These samples were used for validation. Reproducibility of the staining procedure was confirmed using an additional animal (BLI021).

BrdU Treatment

BrdU (Sigma) powder was aliquoted into 20 mg portions in Eppendorf tubes and stored at −20C in the dark. It was dissolved in 1.5 ml saline 5 min before injection. Rats were injected with 50 mg/kg body weight BrdU through the penile vein 1 hr before sacrifice.

Sampling

Samples were randomly collected from each remnant liver lobe: [superior caudate lobe (SCL); inferior caudate lobe (ICL); and right inferior lobe (RIL)] to ensure that every hepatocyte had an equal chance to be analyzed. Tissue samples were sliced into 0.5 × 0.5 × 0.3 cm³ (long, wide, and thick) rectanglular blocks and placed in cassettes. From each lobe, three rectanglular blocks were sampled. A piece of duodenum was added in the same cassette as an internal control for BrdU exposure. The cassette was immersed in 4.5% neutral buffered formalin (Roth; Karlsruhe, Germany) for 24 to 48 hr. Samples were processed, dehydrated, and embedded with paraplast (McComick Scientific; St. Louis, MO). From each block, five sections were cut at 4-μm thick, floating on a water bath at 40C. The sections were mounted on SuperFrost Plus Objekttraeger (R. Langenbrinck; Teningen, Germany) and dried at 37C overnight. One section was stained with hematoxylin-eosin for morphological control. Samples with histopathological changes, such as confluent necrosis or severe sinusoidal dilatation, were excluded. Three sections were used for BrdU staining.

BrdU Staining

Incorporated BrdU was visualized by immunohistochemical staining to allow calculation of the BrdU-LI (Table 1). The staining procedure was based on a modified protocol of Sigma, Inc. After deparaffinization and rehydration (xylene 30 min, 100% ethanol 3 min, 90% ethanol 3 min, 70% ethanol 3 min, distilled water 3 min, TBS for 5 min), tissue sections were treated with prewarmed 0.1% trypsin solution (Sigma) at 37C for 40 min, followed by denaturation of the DNA with 2 N HCl (Merck; Darmstadt, Germany) at 37C for 30 min. In the next step, sections were incubated with 1:50 monoclonal anti-BrdU antibody (Dako; Hamburg, Germany) at 37C for 1 hr, followed by an alkaline-phosphatase–labeled secondary anti-mouse antibody (Immunologic; Duiven, The Netherlands) for 1 hr at room temperature. Color reaction was performed using the Fast Red Substrate System (sensitive) (Dako) for 10 min. The sections were counterstained with Mayer's hemalaun (Merck) for 10 sec, and coverslipped using ImmuMount (Shandon; Pittsburgh, PA).

Table 1.

Standard immunohistochemical protocol for detection of BrdU

Step	Specification	Conditions (time, temperature)
Dewax	Xylol (Shandon; Pittsburgh, PA); ethanol 96%, 80%, 70%; demineralized water ×2	5 min each, RT
Antigen retrieval	Trypsin (Sigma; St. Louis, MO); TBS ×3	40 min, 37C; 5 min each, RT
DNA denaturation	2 N HCl (Merck; Darmstadt, Germany); TBS ×3	30 min, 37C; 5 min each, RT
Blocking	Protein block (Dako; Hamburg, Germany)	1 hr, RT
Primary antibody	Monoclonal anti-BrdU antibody (Dako), 1:50 diluted in antibody diluent (Dako); TBS ×3	1 hr, 37C; 5 min each, RT
Secondary antibody	PowerVision poly-AP-anti-mouse, ready to use (ImmunoLogic; Duiven, The Netherlands); TBS ×3	1 hr, RT; 5 min each, RT
Chromogen	Fast Red substrate system (sensitive) (Dako); demineralized water	10 min, RT; 5 min, RT
Counterstain	Mayer's hemalaun (Merck); running tap water	10 sec, RT; 5 min, RT
Mounting	ImmuMount (Shandon)

BrdU, 5-bromo-2-deoxyuridine; RT, room temperature; ×2, twice; ×3, three times.

Image Acquisition

Images of BrdU-stained sections were acquired at 200× magnification using a microscope (Leica; Wetzlar, Germany). Images were captured using a digital camera (Colorview III; Sony, Japan) connected to both the microscope and a computer (Intel Pentium IV Processor, 2.67 gigahertz, 2-gigabyte DDR main memory, and a 200-gigabyte hard disk) employing a FireWire connection. The software components consisted of an operating system (Microsoft Windows XP), an image analysis package (AnalySIS 5.0, build 1235, Olympus; Watford, UK), and data analysis software (Microsoft Excel XP). The area of each image was 1.42 mm². Images from periportal, pericentral, and parenchymal areas were captured randomly. Areas showing cutting or staining artifacts were excluded from image aquisition. The vessels were located in the middle of the image. The edge of the section and the big vessels (larger than one fourth of the total image area) were excluded from analysis.

Quantification of Cell Proliferation

Manual Counting With Image Tool

Total hepatocytes and BrdU-labeled hepatocytes were counted using images at 200× magnification by five researchers experienced in the analysis of immunohistochemical stainings based on digital photos. Each accepted cellular event was tagged with a mark, and the overlay image was saved.

Automatic Counting With AnalySIS

Digital image analysis starts with the step of color separation (conversion of color images into greyscale images) followed by segmentation (extraction of objects based on a color threshold), parametrization (combination of geometrical properties, characterizing the target object), and particle detection.

Color Separation

After loading of an image into the AnalySIS program, the color image is first converted into greyscale images by separating the color into three basic color channels: blue, green, and red. The total number of hepatocytes was counted using the red channel, based on the dark-blue staining in their nuclei. BrdU-positive hepatocytes were detected using the blue channel, based on the red staining in their nuclei.

Segmentation

The target color channel was activated. The program calculated the relevant gray value range for each image individually. The overflow was set at 49% to ignore 49% of the lightest gray pixels when calculating the thresholds.

Parametrization

The counterstain as well as the antigen-specific detection system (monoclonal anti-BrdU antibody) stained the nuclei not only of the target cells (hepatocytes) but also of other cells, for example lymphocytes, endothelial cells, and Kupffer cells. Therefore, the target objects had to be further characterized by discriminative geometrical properties, which were selected from ∼100 different object characteristics offered by the program (e.g., size and shape). Both mononucleated hepatocytes and binucleated hepatocytes exist in the liver parenchyma. Gandillet et al. (2003) showed that the proportion of binucleated hepatocytes in normal Sprague-Dawley livers was as high as 21%. Nuclei of dinucleated hepatocytes were often found in close proximity, almost appearing as “double nuclei.” Because these “double nuclei” differed in size and shape from the mononucleated hepatocytes, we decided to add additional particle parametrization and detection steps for the binucleated hepatocyte. Ten images (∼20,000 hepatocytes) were selected randomly to determine the mean diameter and sphericity of both mononucleated and binucleated hepatocytes (Figure 1).

Figure 1.

Parameters for parametrization of mononucleated hepatocytes and binucleated hepatocytes.

Particle Detection

Particles fulfilling the criteria of size and shape were detected automatically by AnalySIS. Particles that did not meet the criteria of size and shape were excluded. Results were exported into an Excel worksheet.

Calculation of BrdU-LI

The BrdU-LI was calculated as the percentage of BrdU-labeled nuclei of hepatocytes out of the total number of hepatocytes.

Accuracy Control

An experienced observer reviewed the masked images randomly to check the precision and accuracy of the program.

Validation of Staining Procedure (Based on Experimental Run)

The intra- and inter-experimental variation was assessed by running the same sample (animal identification: BLI021) repeatedly (five times) in one run and by running the same sample in different experiments (five times). The BrdU-LI was obtained by manual counting. Finally, the BrdU-LI, obtained after analysis of 10 images (five pericentral and five periportal areas) per sample, was compared.

Validation of Image Analysis (Based on Defined Images)

To evaluate the computer-based counting with AnalySIS, one section with low proliferation (animal AEE015), two sections with moderate proliferation (animals AEE016 and AEE017), and one section with high proliferation (animal AEE018) were selected randomly. Three images of pericentral, periportal, and parenchymal areas were captured from each section. The results obtained by the conventional counting and computer-based counting using AnalySIS were compared. The overlay images of both counting methods were merged. The numbers of true-positive (TP), true-negative (TN), false-positive (FP), and false-negative (FN) objects were counted in 12 images of the series.

Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated.

The sensitivity was determined with the following formula:

The specificity was determined with the following formula:

The PPV was determined with the following formula:

The NPV was determined with the following formula:

Determination of the Intra-section Variation

To determine the intra-section variation, all pericentral and periportal areas of one section (animal MAR001) were captured and analyzed. The BrdU-LIs obtained for each image were compared.

Determination of the Intra-block Variation

To determine the intra-block variation, three blocks (blocks 215, 219, and 222) were obtained from SCL, ICL, and RIL lobes, and three sections were cut.

All three sections from each block were stained. Five pericentral and five periportal areas were captured and analyzed. The BrdU-LIs were compared.

Determination of the Intra-lobe Variation

To determine the intra-lobe variation, tissue was sampled from three remnant lobes (animals MAR001; SCL, ICL, and RIL). Three blocks from each lobe were sampled and cut, and one section from each block was stained. Five pericentral areas and five periportal areas from each section were captured and analyzed. The BrdU-LIs were compared.

Determination of the Inter-lobe Variation

To determine the inter-lobe variation, tissue was sampled from the three remnant lobes of the same animal (MAR001; SCL, ICL, and RIL). Five sections from each lobe were cut and stained. Five pericentral areas and five periportal areas were captured and analyzed for each section. The BrdU-LIs were compared.

Estimation of the Total Size of the ROI to Be Analyzed

The total size of the ROI to be analyzed was determined by both the classical statistical approach based on the central limit theorem and the lobule-dependent zonal measurement (LZM) method (Bahnemann and Mellert 1997). A coefficient of variation (CV) of below 10% was considered to be acceptable, inasmuch as the biological variation between individual animals was much higher and can reach more than 30%, as already shown by Fabrikant as early as 1968 (Fabrikant 1968;Schmitz and Hof 2000).

Based on the central limit theorem, the sample size of ROI (n_t) was calculated using the following formula:

SD_t: the standard deviation of whole liver; t_{α, n−1}: the 95th percentile (α = 0.05) of a t distribution with n−1 degrees of freedom (n=total number of ROIs); d: the desirable difference in absolute value between the observed mean and the true mean.

The total size of the ROI to be analyzed, consisting of sample size of the number of pericentral n_z and periportal n_p areas, was calculated as follows:

n_z: the sample size of pericentral area; n_p: the sample size of periportal area; n_zt: the total number of pericentral areas; n_pt: the total number of periportal areas; SD_z: the standard deviation of all the pericentral areas; SD_p: the standard deviation of all the periportal areas.

Based on the LZM method, the minimum number of events to be analyzed to obtain the mean BrdU-LI representing the “true mean” with a CV below 10% was determined. Five observers analyzed the same section (animal MAR001: section 215a) and used different numbers of ROIs as follows: 1 pericentral and 1 periportal area (Z1/P1), 3 pericentral and 3 periportal areas (Z3/P3), 5 pericentral and 5 periportal areas (Z5/P5), 10 pericentral and 10 periportal areas (Z10/P10), and 20 pericentral and 20 periportal areas (Z20/P20). ROIs were selected according to the rules of the LZM method. The minimum number of ROIs was determined by comparing the results obtained from the five analyzers.

Statistics

The data were analyzed using Sigmaplot 10.0 in combination with SigmaStat 3.0 (Statcon; Witzenhausen, Germany). Differences between paired groups were analyzed using the two-tailed paired samples Student's t-test, and differences between independent groups were analyzed using the two-tailed independent samples Student's t-test. Multiple groups were compared using the one-way independent ANOVA test. The α level was defined as 0.05. The degree of variation within images, sections, blocks, and lobes of the samples was described by calculating the CV, (CV = standard deviation/mean).

Results

Validation of Staining Procedure (Intra- and Inter-assay Variability)

Results obtained from different immunohistochemical runs can only be compared if it is ascertained that there is no intra- and inter-experimental difference. Therefore, the same sample was run repeatedly in the same assay and was later used as control in the following assays. Repeated staining and evaluation of the same sample within the same assay (intra-assay reproducibility) did not lead to a statistically significant difference between the results (ANOVA = 0.962) (data not shown). Moreover, the CV was as low as 7.57%. Sections of the same sample used as controls in the subsequent experiment runs underwent quantitative analysis (n=5×). Again, the results were slightly different but did not reach statistical significance (ANOVA = 0.889) with a CV of below 10% (9.69%) (data not shown). This indicated that our staining result was highly reproducible within and between experiments.

Validation of Computer-based Counting

Experienced observers, specifically trained in the quantification of stained hepatocyte nuclei, analyzed a series of images repeatedly to exclude any observer-related bias. Results varied in terms of single cells but did not show any statistically significant differences for individual observers or between observers.

To validate the computer-based counting, the overlay images of both computer-based counting and conventional counting were merged. The numbers of TP, TN, FP, and FN objects were counted, and sensitivity, specificity, PPV, and NPV were calculated to evaluate the computer-based counting. For the total number of hepatocytes, the sensitivity was 94.39%, specificity was 79.12%, PPV was 82.99%, and NPV 92.92% (Table 2). For BrdU-labeled hepatocytes, sensitivity was 91.38%, specificity 99.45%, PPV 97.64%, and NPV 99.13% (Table 3). The result generated with the computer-based counting was well correlated with the result from manual counting. The correlation coefficient (r) was 0.9976 (Figure 2). It suggested that the result from the computer-assisted counting was valid and correlated highly with the result from manual counting.

Table 2.

Sensitivity and specificity of the computer-assisted image analysis in counting total hepatocytes

Image	TP	FP	FN	TN	Sensitivity	Specifity	PPV	NPV
AEE15-T07-103a	185	38	17	136	91.58	78.16	82.96	88.89
AEE15-1-T07-103b	200	45	20	182	90.91	80.18	81.63	90.10
AEE15-1-T07-103	182	51	5	155	97.33	75.24	78.11	96.88
AEE16-T07-106a	189	37	10	131	94.97	77.98	83.63	92.91
AEE16-T07-106b	176	25	18	127	90.72	83.55	87.56	87.59
AEE-017-T07-116a	209	35	13	144	94.14	80.45	85.66	91.72
AEE-017-T07-116b	175	50	18	151	90.67	75.12	77.78	89.35
AEE-017-T07-116	201	22	1	147	99.50	86.98	90.13	99.32
AEE-18-4-T07-119a	182	59	9	99	95.29	62.66	75.52	91.67
AEE-18-4-T07-119b	177	46	10	227	94.65	83.15	79.37	95.78
AEE-18-4-T07-119	200	21	3	139	98.52	86.88	90.50	97.89
Mean					94.39	79.12	82.99	92.92
SD					3.17	6.82	5.07	3.97

^aPericentral area.

^bPeriportal area.

TP, true positive; FP, false positive; FN, false negative; TN, true negative; PPV, positive predictive value; NPV, negative predictive value; SD, standard deviation.

Table 3.

Sensitivity and specificity of the computer assisted image analysis in counting BrdU-labeled hepatocytes

Image	TP	FP	FN	TN	Sensitivity	Specifity	PPV	NPV
AEE15-T07-103a	1	0	0	375	100.00	100.00	100.00	100.00
AEE15-1-T07-103b	2	0	0	445	100.00	100.00	100.00	100.00
AEE15-1-T07-103	1	0	1	391	50.00	100.00	100.00	99.74
AEE16-T07-106a	56	2	6	303	90.32	99.34	96.55	98.06
AEE16-T07-106b	69	0	8	269	89.61	100.00	100.00	97.11
AEE-017-T07-116a	20	0	1	380	95.24	100.00	100.00	99.74
AEE-017-T07-116b	22	0	2	370	91.67	100.00	100.00	99.46
AEE-017-T07-116	37	2	0	332	100.00	99.40	94.87	100.00
AEE-18-4-T07-119a	49	3	2	295	96.08	98.99	94.23	99.33
AEE-18-4-T07-119b	70	5	2	383	97.22	98.71	93.33	99.48
AEE-18-4-T07-119	115	6	6	236	95.04	97.52	95.04	97.52
Mean					91.38	99.45	97.64	99.13
SD					14.23	0.79	2.81	1.05

^aPericentral area.

^bPeriportal area.

Figure 2.

The 5-bromo-2-deoxyuridine labeling index (BrdU-LI) generated by computer-based automatic image analysis compared with those generated manually by a skilled image analyzer. A strong correlation between the two sets of results was obtained (r = 0.9976).

Analysis Time

For the image-based manual counting, all observers were trained in the histological identification of liver structures by the experienced pathologist. It normally took ∼1 week to train an unexperienced observer to reach a sensitivity, specificity, accuracy, and precision of more than 90%. Manual analysis using Image tool needed ∼8 min per image when the BrdU-LI was below 10% and ∼11 min per image when the BrdU-LI was ∼30% (performed by experienced observer MD). In contrast, operators needed only brief instruction (5–10 min) to be able to operate the computer-based image analysis system. Because the image analysis system was fully automatic, even an observer not trained in histology could run it and obtain the same results as the experienced observer. Running the AnalySIS-Macro took only 1 min per image. The experienced observer only had to review the masked image randomly to control the precision and accuracy of the program. Therefore, the computer-assisted automatic counting could reduce the analysis time of BrdU-LIs by more than 90%.

Intra-section Variation

The BrdU-labeled hepatocytes were distributed inhomogeneously throughout the whole section (Figure 3). The BrdU-positive hepatocytes were located mainly in the periportal areas. Hence, in the pericentral areas, the number of BrdU-positive hepatocytes was much lower than in the periportal areas. This indicated that liver regeneration was inhomogeneous in different liver zones.

Figure 3.

Distribution of BrdU-labeled hepatocytes. More BrdU-labeled hepatocytes were distributed in the periportal area than in the pericentral area (animal MAR001). PT, portal trial; ZV, central vein; HPC, hepatocyte.

To assess the intra-section variation, all pericentral areas and all periportal areas of one randomly selected section (animal MAR001: section 215a) were captured and analyzed. The range of BrdU-LIs within the pericentral areas (n=26) varied from 6.05% to 31.17% (Figure 4). The range of BrdU-LIs within the periportal areas (n=34) varied from 29.95% to 54.29%. The difference in the BrdU-LIs between the pericentral (19.12 ± 5.29%) and the periportal (40.04 ± 6.34%) areas even reached statistical significance (p<0.001). Taken together, the BrdU-LIs of all images obtained from the whole section ranged from 6.05% to 54.29%, resulting in a mean of 30.66% and a large standard deviation of 12%. The CV of the whole section was as high as 39.17%, which suggested that the intra-section variation was large. Therefore, the selection of ROIs by different observers can influence the result of the BrdU-LI, especially when looking at a small number of ROIs, such as a single image representing only one portal field or one central vein area.

Figure 4.

BrdU-LI of all the pericentral and periportal areas within the same section (animal MAR001; section 215a). Data are shown as mean ± SD. PT, periportal area; ZV, pericentral area; **p<0.01 vs ZV.

Intra-block Variation

Sections were cut at 4-μm thickness, which is less than the diameter of a hepatocyte nucleus and obviously thinner than the hepatocyte itself. Depending on the cutting level, the nucleus of a hepatocyte may not be visible in a given section and therefore may escape from an analysis based on a single two-dimensional section. Because this systematic error might affect the result of the quantitative assessment of hepatocyte proliferation within a given block, we examined serial sections from the same block.

To determine the extent of intra-block variation, three sections from three blocks of each lobe (SCL, ICL, and RIL) were stained. Five pericentral and five periportal areas per section were analyzed. The mean BrdU-LI was 28.93 ± 1.09% in the SCL, 30.68 ± 2.51% in the ICL, and 39.00 ± 1.34% in the RIL. The CV of the three blocks ranged from 3.44% to 8.21% (Table 4), all lower than our attempted CV of below 10%. These results indicated that the intra-block variation might not be a key factor influencing the result of the BrdU-LI.

Table 4.

Intra-block variation of BrdU-LI

Lobe		SCL			ICL			RIL
Block		215			219			222
Section	215a	215b	215c	219a	219b	219c	222a	222b	222c
Mean of section	28.79	27.92	30.09	31.94	27.77	32.31	37.48	40.02	39.49
Mean of block		28.93			30.68			39.00
SD of block		1.09			2.52			1.34
CV of block (%)		3.77			8.21			3.44

BrdU-LI, 5-bromo-2-deoxyuridine labeling index; SCL, superior caudate lobe; ICL, inferior caudate lobe; RIL, right inferior lobe; CV, coefficient of variation.

Intra- and Inter-lobe Variation

Subjecting the liver to a surgical procedure might influence the perfusion of the liver. Alterations in hepatic microcirculation and perfusion, especially after a major proliferation stimulus is exerted on the liver, as done with PH, might affect liver regeneration (Dirsch et al. 2008). Therefore, standard experimental hepatectomy might influence hepatic microcirculation and perfusion. Subsequently the spatial distribution of hepatocyte proliferation within and between the remnant liver lobes may follow a different kinetic, which would be reflected in a highly variable zonally dependent hepatocyte BrdU-LI within and between liver lobes.

To determine the intra-lobe variation, three samples from each remnant liver lobe were collected. One section from each sample was stained. Five pericentral areas and five periportal areas per section were analyzed. The mean BrdU-LI was 28.10 ± 2.77% in the SCL, 31.51 ± 2.42% in the ICL, and 39.29 ± 1.05% the RIL. The CV of the three lobes ranged from 2.68% to 9.84% (Table 5). This suggested that the intra-lobe variation might not be the key factor influencing the result of the BrdU-LI, when basing the assessment on a total ROI of five periportal and five pericentral images.

Table 5.

Intra- and inter-lobe variations of BrdU-LI

Lobe	SCL					ICL					RIL
Section	215a	215b	215c	216	217	218	219a	219b	219c	220	222a	222b	222c	223	224
PT1	34	38	39	21	38	29	33	31	38	33	45	44	40	41	51
PT2	20	39	39	26	19	35	37	25	32	40	53	50	52	52	57
PT3	31	32	44	40	19	49	45	41	50	42	50	48	45	53	48
PT4	40	36	45	26	47	47	36	39	42	55	47	58	44	52	53
PT5	38	36	44	27	37	46	53	29	45	46	55	46	56	53	57
ZV1	33	21	25	14	20	31	20	20	22	32	24	22	35	35	23
ZV2	24	21	14	17	19	24	25	23	14	11	21	35	29	31	29
ZV3	20	19	19	25	26	31	23	31	37	14	21	38	35	34	26
ZV4	23	18	16	17	43	32	15	22	18	20	21	30	27	26	23
ZV5	25	19	16	22	35	20	34	18	25	18	36	29	32	24	26
Mean of section	29	28	30	23	30	34	32	28	32	31	37	40	39	40	39
Mean of lobe	28.10					31.51					39.29
SD of lobe	2.77					2.42					1.05
CV of lobe (%)	9.84					7.69					2.68
Mean of liver	32.96
SD of liver	11.94
CV of liver (%)	36.23

ZV, pericentral area; PT, periportal area.

To determine the inter-lobe variation, five sections from each lobe (SCL, ICL, and RIL) were stained. Five pericentral areas and five periportal areas per section were analyzed. The mean BrdU-LI of the RIL was 39.29 ± 1.05%, whereas the mean BrdU-LI of the SCL was only 28.10 ± 2.77%, which was more than 10% lower than the BrdU-LI of the RIL (Table 5). That suggested that spatial distribution of the average periportal/pericentral hepatocyte proliferation was highly variable in the comparison of the three different hepatic lobes, whereas the variation was lower within a given lobe. The results confirm the necessity of keeping the sampling procedure standardized, especially always selecting the same remnant liver lobe for quantifying liver regeneration in terms of the BrdU-LI after surgical reduction of the liver mass.

Estimation of the Total ROI to Be Analyzed

Assessment of liver regeneration based on different liver lobes and several blocks from the same lobe demonstrated the highly variable spatial distribution of hepatocyte proliferation with a BrdU-LI of 32.96 ± 11.94%, as indicated in Table 5. Therefore, it is necessary to estimate the appropriate size of the ROI to be analyzed to ensure that the observed mean represents the true mean. The size of the ROI was estimated with both the classical statistical method and the LZM method based on the above result.

Classical Statistical Method (Central Limit Theorem)

As demonstrated in Table 5, the mean BrdU-LI of 150 images in total, representing the three remnant liver lobes, was 32.96 ± 11.94%. Based on a calculated CV of ∼10% when attempting to reach a maximal difference of 3% in absolute value between the observed mean and the true mean, 43 ROIs (18 pericentral and 25 periportal areas) were calculated to be necessary for analysis of the whole remnant liver (α = 0.05, n=150). Because the inter-lobe variation was a key variable, the total size of the ROIs to be analyzed in a given lobe was calculated. According to the standard deviation of each lobe (Table 5), the total size of the ROIs to be analyzed was calculated based on central limit theorem. The total size of the ROIs to be analyzed was calculated as follows: 19 ROIs for SCl, 16 ROIs for ICL, and 8 ROIs for RIL.

LZM Method

According to the rules of LZM, all pericentral and periportal areas from one section were analyzed. Five observers analyzed different numbers of randomly selected ROIs as follows: Z1/P1, Z3/P3, Z5/P5, Z10/P10, and Z20/P20.

The mean BrdU-LI based on P1/Z1 varied between 18.89% and 40.46%, depending on the observer (Table 6). The CV between the results generated by the five different observers was as high as 34.36%, reflecting the spatial distribution of proliferative “hot spots.” This effect leveled off when the total size of ROIs to be analyzed was increased. When P3/Z3 was analyzed, the BrdU-LI ranged from 23.09% to 39.3%, with a CV between the five observers of 27.01%. In the analysis of P5/Z5 or more, the BrdU-LI only ranged from 27.15% to 30.68%, with a low standard deviation of 1.66 and a CV of 7.43%, which was in the desired range of below 10%.

Table 6.

BrdU-LI generated with different “total size of ROI to be analyzed”

	P1/Z1	P3/Z3	P5/Z5	P10/Z10	P20/Z20
Analyzer 1	18.89	27.26	27.15	30.55	28.24
Analyzer 2	40.46	33.68	28.23	28.77	28.17
Analyzer 3	29.65	39.30	30.68	28.15	28.42
Analyzer 4	30.75	23.09	27.15	32.10	30.29
Analyzer 5	26.64	28.22	30.13	28.63	29.43
Mean	29.28	30.31	28.67	29.64	28.92
SD	7.78	6.28	1.66	1.65	0.92
CV (%)	34.36	27.01	7.43	7.22	4.11

Z, pericentral area; P, periportal area.

In conclusion, the intra-section variation (image-to-image, potentially reflecting microcirculatory differences) and the inter-lobe variation (lobe-to-lobe, potentially reflecting perfusion differences) were the key variables influencing the BrdU-LI in a remnant liver. Therefore, one section from each lobe should be quantified to obtain the true mean of the remnant liver, and one lobe should be designated as representative lobe for a given experiment.

Discussion

The more we understand the molecular basis of disease and the more drugs with known molecular targets enter the market, the more important becomes quantification of results in molecular pathology as a basis of diagnostic and therapeutic decisions. The quantification has to be of high accuracy and reproducibility, and the procedure has to be time efficient and thereby affordable.

The main reason for the present study emanated from the need for a reliable, robust, fast, and cost-efficient method for assessing the proliferative index of hepatocytes in regenerating livers by using the BrdU-LI.

To attain this goal, we first tried to achieve consistent and reproducible staining results. We applied standardized preanalytical and analytical procedures following the guidelines issued by Ruehl-Fehlert et al. (2003) and Nolte et al. (2005). Both inter-assay and intra-assay results showed no statistically significant differences and a CV of below 10% in serial sections. Thus, strict adherence to defined standards in tissue handling and the immunohistochemical procedures allowed the generation of highly reproducible results in our setting.

The quantification of the BrdU-LI by manual counting is highly time-consuming. Furthermore, the results are dependent on the experience and the bias of the individual observer. Therefore, we generated and validated a fully automatic AnalySIS-Macro to replace this manual step. Soames et al. (1994) also reported an image analysis system for the quantification of the BrdU-LI. We compared our system with their system in the following aspects: color separation, parameter for single nuclei, and parameters for binucleated hepatocytes. In their system, counterstain was not applied prior to counting of BrdU-labeled hepatocytes to maximize the contrast for the labeled nuclei. This suggested that they could not use the same image for counting both BrdU-labeled and non-labeled nuclei. Our system can transfer the color image into gray images by separating the red, green, and blue color. By choosing defined color channels for defined color staining (e.g., choosing blue channel for red-stained nuclei), the contrast between nuclei and background was increased, so that the nuclei could be clearly delineated and subjected to segmentation. Therefore, the same image could be used for counting both BrdU-labeled and total hepatocytes automatically with our system. The proportion of binucleated hepatocyte was over 20% in our samples. It was necessary to create an algorithm suitable for the detection of these binucleated hepatocytes. Soames et al. (1994) did not describe a step to detect binucleated hepatocyte but claimed that their system identified 90% of true hepatocytes. The sensitivity of our system was 94.39% in total hepatocytes and 91.38% in BrdU-labeled hepatocytes. Furthermore, the BrdU-LI generated in the conventional manner and in the computer-based manner was more strongly correlated in our system (r=0.99) than in the Soames system (r=0.92). In comparison to Soames' system, our system led to more-accurate results.

Because the system is fully automatic, the result is fully reproducible and completely observer independent, when using the same images. It eliminates the observer-related bias in identifying the target events. Furthermore, it reduces the analysis time by 90%, and thereby greatly improves the economical efficiency.

The aim of a scientific sampling regime is that every cell should have equal chance to be analyzed in a given biological system, such as a whole liver, a given lobe, and a given part. This equal probability guarantees that a study is representative for a specific biological system. Therefore, random collection of tissue samples in every liver lobe is important to eliminate a potential sampling bias.

The use of a fully automated image analysis system can exclude the bias from individual observers. However, we are aware that a two-dimensional image–based analysis system could not avoid the bias in size selection of the target particles. Large hepatocytes such as binucleated hepatocytes have a higher probability of appearing at a randomly selected thin tissue section compared with smaller ones such as mononucleated hepatocytes. Furthermore, the size of the hepatocyte appearing larger or smaller is highly dependent on the cutting level. The same hepatocyte nucleus appears rather large when the tissue slice is obtained from the middle of the nucleus compared with the top or bottom of the ball-shaped nucleus. This may lead to a relative underestimation of the number of small hepatocytes and may finally cause an error in the estimation of the total number of hepatocytes. However, the classification criteria, which are applied for the selection of BrdU non-labeled hepatocytes are the same as those applied for the selection of BrdU-labeled hepatocytes. The BrdU-LI represents the percentage of BrdU labeled in non-labeled and labeled hepatocytes. Therefore, the true number of total hepatocytes within an image may be less important for the quantification of the BrdU-LI than for the estimation of the number of total hepatocytes within a liver.

With the development of the disector principle, design-based stereology can accurately and precisely estimate the number of total events not relying on size, shape, or orientation (Gundersen et al. 1999; Mühlfeld et al. in press). Marcos et al. (2006) used a design-based stereological approach to estimate the total number of hepatocytes in a rat liver. In this study, 3565 hepatocytes per liver were counted manually to estimate the total number of hepatocytes with a CV of below 10%. Manual counting of such a high number of hepatocytes represents a huge workload, leading to a low economical efficiency.

In contrast, our sample size was much larger than that in the system of Marcos (8000 hepatocytes for a whole remnant liver or 2000 hepatocytes for one lobe vs 3565 hepatocytes for a whole liver), but the workload demand was much lower.

In summary, our approach may reduce the size-related bias in our two-dimensional–based analysis, but cannot fully eliminate it. Observer-related and size-related bias can only be excluded by combining a computer-based automated three-dimensional image analysis system with a design-based stereological approach for quantification of hepatocytes, respectively, by analyzing all serial sections of the remnant liver.

Our image analysis system can reduce the observer-dependent bias and thereby reduce part of the postanalytical variation. Because liver regeneration is inhomogeneous within a liver lobe, but also between liver lobes, the selection and the total size of the ROIs to be analyzed can also influence the result of the BrdU-LI on the postanalytical level.

Bahnemann and Mellert (1997) described the LZM method for determination of cell proliferation in the liver. According to the rules of the LZM method, measurement fields should be of equal size. In the present study, we defined the volume of all liver samples as a 0.5 × 0.5× 0.3 cm³ rectanglular block. Furthermore, all images were captured at 200× magnification, so that the surface area of all measurement fields was identical. We would like to recommend such a standardized procedure to ensure a total ROI of equal size.

According to the LZM method, a periportal and a pericentral area should be included and zonal division should be performed. The rough division of the lobule into three zones is only an approximation of the actually existing physiological differences (Thung and Gerber 1991). It is difficult to strictly perform the zone division, especially when the liver architecture is impaired (Wagenaar et al. 1993). Because the difference in the proliferative response between the pericentral and periportal areas is widely accepted (Gebhardt and Jonitza 1991; Chen et al. 1995), we recommend dividing the measurement fields clearly into pericentral and periportal areas instead of different zones.

According to the LZM method, measurement fields should be distributed equally over the lobe(s), excluding the edge of the lobe and excluding large vessels. Our determination of diverse variations in the BrdU-LI demonstrated that the intra-block and the intra-lobe variations were small (CV <10%). They might not be the key variables influencing the result of the BrdU-LI. This suggested that the BrdU-LI of a randomly selected section might represent the BrdU-LI of the respective lobe. Nevertheless, our results demonstrated that the mean of the BrdU-LIs from different lobes varied from 28.1% (SCL) to 39.29% (RIL) within the same liver, a range clearly exceeding 10%. This difference in proliferation between liver lobes might be attributed to differences in hepatic perfusion subsequent to the surgical procedure. In contrast to our method, a toxicological liver injury model was used in the study of Bahnemann and Mellert (1997). The variations in treatment and liver perfusion throughout the whole liver might be minor, and therefore the inter-lobe variation in the BrdU-LI might also be minor.

However, these observations require some considerations. First, liver regeneration in different liver lobes may be seriously affected by the experimental surgical procedure itself and not only by the experimentally selected variable such as a drug. Second, because the inter-lobe variation of the BrdU-LI may be large, the BrdU-LI of samples from different lobes of different individuals may not be comparable, especially when using surgical models as the regeneration stimulus. Third, interpretations of differences in the BrdU-LI should take these observations into consideration. Because the BrdU-LI of the same liver can differ by more than 10% when selecting different liver lobes, the BrdU-LI result of different experimental groups may be statistically significantly different, but not necessarily related to the experimental condition or biologically relevant.

Therefore, we recommend that the selection of ROIs to be analyzed should be restricted to a defined liver lobe throughout a given experiment, or they should be distributed in all lobes equally.

According to the LZM method, ∼1000 hepatocytes per zone should be examined. In other studies, counting as few as three fields or 1000 hepatocytes was considered to be reliable (Soames et al. 1994; Urade et al. 1996). In our study, 43 ROIs (8000 hepatocytes) per liver were calculated to be necessary to reach the level of 3% difference from the true mean when using the classical statistical method. Such a large number of ROIs is less strenuous when using an image-based automatic counting procedure. However, analyzing so many ROIs manually for a given experiment is a high and extremely time-consuming workload. Therefore, the number of ROIs to be analyzed was also estimated based on the LZM method and our results. Our results suggested that 5Z/5P (∼2000 hepatocytes) were sufficient to reach a mean that was representative of the true mean of the BrdU-LIs for a given section within a CV below 10%. Analyzing a smaller total number of ROIs by using 1Z/1P and 3Z/3P was associated with a high CV ranging from 20% to 40%. Therefore, we recommend that at least 2000 hepatocytes per lobe should be examined to achieve the representative result of this given lobe, and that the number of pericentral areas should be equal to the number of periportal areas.

Taken together, standardization of the selection and the number of ROIs is crucial to achieving a representative result for the BrdU-LI that is close to the true mean. We recommend standardizing the selection and number of ROIs based on the following: ROIs should have an equal surface; they should be divided into pericentral and periportal areas, and the number of pericentral areas should be equal to periportal areas; ROIs should either be restricted to a defined liver lobe throughout the whole experiment, or they should be distributed in all lobes equally; and at least 2000 hepatocytes per lobe should be examined.

Summary and Conclusion

Our newly developed and validated automatic AnalySIS-Macro proved to be a highly economical, efficient, and fully automated tool to detect BrdU-labeled and non-labeled hepatocyte nuclei with high sensitivity and specificity. It turned out to be a useful tool for fast, reliable, and reproducible quantification of proliferating hepatocytes in regenerating livers. The use of this tool in combination with rules for the selection and size of ROIs can be considered a prerequisite for assessing the spatial distribution of cellular events such as proliferating hepatocytes, inasmuch as it allows the analysis of large ROIs in a reasonable amount of time.