Automated biphasic morphological assessment of hepatitis B-related liver fibrosis using second harmonic generation microscopy

Liver fibrosis assessment by biopsy and conventional staining scores is based on histopathological criteria. Variations in sample preparation and the use of semi-quantitative histopathological methods commonly result in discrepancies between medical centers. Thus, minor changes in liver fibrosis might be overlooked in multi-center clinical trials, leading to statistically non-significant data. Here, we developed a computer-assisted, fully automated, staining-free method for hepatitis B-related liver fibrosis assessment. In total, 175 liver biopsies were divided into training (n = 105) and verification (n = 70) cohorts. Collagen was observed using second harmonic generation (SHG) microscopy without prior staining, and hepatocyte morphology was recorded using two-photon excitation fluorescence (TPEF) microscopy. The training cohort was utilized to establish a quantification algorithm. Eleven of 19 computer-recognizable SHG/TPEF microscopic morphological features were significantly correlated with the ISHAK fibrosis stages (P < 0.001). A biphasic scoring method was applied, combining support vector machine and multivariate generalized linear models to assess the early and late stages of fibrosis, respectively, based on these parameters. The verification cohort was used to verify the scoring method, and the area under the receiver operating characteristic curve was >0.82 for liver cirrhosis detection. Since no subjective gradings are needed, interobserver discrepancies could be avoided using this fully automated method.


Results
Advanced SHG/TPEF microscopic quantification of fibrillar collagen was highly correlated with ISHAK fibrosis scores. To elucidate the SHG/TPEF microscopy imaging capacity for liver fibrosis, 175 liver biopsy samples were obtained from chronic HBV-infected patients. The clinical parameters of the recruited patients are listed in Table 1. One hundred and five randomly selected samples were assigned to the training cohort and used for developing and optimizing the computer-based algorithm for liver fibrosis quantification. The other 70 samples were used as a validation set to assess the performances of the models. Figure 1A-D shows the images acquired using the conventional staining (Fig. 1A,B) and SHG/TPEF microscopy (Fig. 1C,D) methods. In SHG/TPEF imaging, fibrillar collagen was detected by SHG and is shown in green, whereas hepatocyte morphology was captured by TPEF microscopy and is shown in red. Visually, the SHG/TPEF technology resulted in the generation of a much more enhanced image with a better collagen distribution contrast. To minimize interference with normal-structured parenchyma and portal tract collagen, normal collagen readings were excluded. Figure 1E depicts the correlations identified by basic quantification (total collagen) or advanced quantification (Fibro-C-Index) 27 , excluding the normal-structured collagen signals 31 , with the ISHAK scores. The advanced quantification results were highly correlated with the ISHAK fibrosis scores.
Scientific RepoRts | 5:12962 | DOi: 10.1038/srep12962 Morphological liver fibrosis features visualized by SHG/TPEF microscopy. With the SHG/ TPEF system, liver fibrosis severity could be quantitatively measured by examining the collagen fiber amount and morphology. Figure 2 shows the collagen profiles obtained from liver fibrosis qualitative imaging of the liver biopsies, including the total collagen distribution (A, green), various collagen strings (B, colored region), collagen percentage of the total area (CPA) (C), and collagen crosslinks (D, white arrow). The SHG/TPEF system employed 19 features (Table 2) to quantitatively measure liver fibrosis severity. These features were classified into three categories. The first feature category (F1-3) was collagen percentage, which included the total, aggregated and distributed collagen percentages. The second feature category (F4-13) was based on collagen string detection. Once a collagen string was detected, the contour that enclosed the collagen pixels was defined. Subsequently, the convex hull and the area and perimeter of each collagen string contour were also measured (Fig. 3A). To further describe the shapes and properties of the detected collagen strings, for each string, an ellipse was applied that enclosed  Table 1. Basic clinical data for the patients whose liver tissues were submitted for fibrosis evaluation. all of the pixels used to calculate the global parameters (e.g., breadth and length) (Fig. 3B). Collagen connectivity (Fig. 3C) within the strings was calculated by skeletonizing them and then detecting the intersections between them. The last set of features (F14-19) was generated according to the ratios of the different collagen string types.
Correlation between the morphological features observed with SHG/TPEF microscopy and ISHAK fibrosis stages. To identify the critical morphological features that can be used to assess liver fibrosis, the correlations between the 19 morphological features and ISHAK fibrosis scores were analyzed using univariate linear regression. Assuming that the ISHAK fibrosis scores were a linear function that increased from 1 to 6, we calculated the P values of all 19 features. The results are shown in Table 2. Note that most of the first and second category features (11 of 13 features) were highly correlated (P < 0.001), whereas only some of the third category features were correlated (P < 0.05).
Biphasic algorithm development for liver fibrosis assessment. In statistical analysis, collagen percentages (F1-3) were first employed to perform model fitting according to the ISHAK fibrosis scores (scores 1-6). Second, the collagen string properties (F4-13) were added. Finally, the collagen string ratios (F14-19) were included. A receiver operating characteristic (ROC) curve was used to assess the performance of the SHG/TPEF fibrosis score, using samples from the verification cohort (Table 3). Overall, the sophisticated non-linear SVM model was superior to the generalized linear model. In both models, the additional morphological features significantly increased the area under the ROC curve (AUC). Additionally, some features were more important for identifying different stages, e.g., the F4-13 string properties in the SVM model were used to identify early-stage progression, and the F14-19 string ratios in the generalized linear model were used to identify late-stage progression.
Most combinations possessed high AUC values (most were higher than 0.7, but some were higher than 0.8), indicating that these features were crucial indices that were highly compatible with the clinically accepted ISHAK scores (i.e., these SHG-detected morphological features might facilitate efficient liver fibrosis assessments).
Among all of the models shown in Table 3, the largest AUC was measured while distinguishing between the non-cirrhotic (ISHAK stages 1-4) and cirrhotic (ISHAK 5-6) patients with the multivariate generalized linear model, which utilized all 19 included features (AUC = 0.829). Additionally,

Discussion
Liver biopsy remains the standard for monitoring fibrosis progression, in which a small piece of liver tissue is removed using a biopsy needle, stained, examined under a microscope, and graded based on a descriptive or semi-quantitative scoring method 32 . In addition to the potential complications related to the biopsy procedure, inherent drawbacks exist, including potentials for sampling errors, staining variations, and inter-and intra-observer variabilities, in the interpretation of histology results 32,33 .
With the development of both mode-locked lasers and highly sensitive optical sensors, non-linear optical microscopy, such as multi-photon excitation fluorescence and multi-harmonic generation, has become an affordable option for tissue imaging 34,35 . Because TPEF microscopy has a different excitation mechanism than SHG, TPEF signals can be easily separated from SHG signals using the appropriate detectors. Combined TPEF/SHG images reveal additional fibrotic liver data, such as precise collagen distributions and undistorted liver cell morphologies. In recent studies, TPEF microscopy has been widely used for imaging organ structure components and their dynamic interactions in biological tissues. Additionally, SHG microscopy has been increasingly used for measuring highly ordered structures without central symmetry, such as type I collagen, which is the dominant collagen type in fibrotic livers [36][37][38] .
Compared with conventional collagen imaging, which uses transmitted light microscopy of histological tissue sections that are stained with either Masson's trichrome or picrosirius red, SHG microscopy has several advantages, as follows: (1) this method provides superior information regarding fibrillar collagen; (2) sample staining is unnecessary; thus, the staining variations that result from different stain batches, protocols, time-dependent fading, and photobleaching 26,39 are eliminated; (3) fluorophores are absent from tissues; thus, signals are unaffected by dye concentrations and photobleaching; (4) infrared excitation sources can be used, resulting in less scattering in the tissue; and finally, (5) SHG microscopy enables deeper tissue penetration for imaging purposes; hence, the 3D visualization of fiber architecture can be achieved 40 . Additionally, TPEF microscopy can provide complementary structural and cellular information 41 .
Fibrillar collagen formation occurs late in fibrogenesis. A traditional fibrosis microscopy evaluation is commonly based on restricted assessments related to later fibrotic changes, whereas features of the early stages of fibrogenesis are usually neglected. However, the use of new morphological criteria in conjunction with standard histopathological fibrosis scoring could improve the assessments of relatively small amounts of fibrillar collagen within fibrous tissues. This study evaluated 19 morphological features of SHG/TPEF images that are automatically included by the Genesis program. Morphological features such as collagen string formation and connectivity may allow for the identification of several stages in the progression of fibrillar collagen formation; therefore, these morphological features may enable the evaluation of early stages of fibrogenesis. Gailhouste et al. 42 have highlighted the benefits of the linear scale provided by the SHG index by comparing it with the histological METAVIR scoring system. However, they have shown that the continuous scale provides greater accuracy than the semi-quantitative scoring system because it underestimates variables that demonstrate architectural differences in collagen, resulting in the restricted SHG indexing of fibrosis stages.
In this study, a biphasic model was developed to separately assess the early and late stages of fibrosis. We compared a multivariate generalized linear model with the SVM model. The results showed that the SVM approach was superior for detecting and differentiating between fibrosis stages (ISHAK 1-4),  suggesting that collagen accumulation during fibrosis progression might not be a linear process. The finding that the biphasic approach displayed better accuracy than the monophasic approach further suggests that the pace of collagen accumulation differs during the early and late stages of fibrosis development. After incorporating the SVM into the biphasic approach, we successfully achieved an ROC of > 0.82 for monitoring late liver fibrosis stages (cirrhosis) and an ROC of > 0.75 for monitoring early liver fibrosis stages. We also observed that the addition of many parameters with relatively high P values (P > 0.001) resulted in better performance for assessing the fibrosis degree, especially for the non-linear models. For example, F14-19, which includes the different collagen string feature ratios, was not directly correlated with the degree of fibrosis; however, the addition of these parameters to the multivariate generalized linear model resulted in significant improvements in the ROC values in most cases. AUC analysis revealed that the inclusion of some features, such as the string properties (F4-13), into the SVM models helped to improve the distinction of early stage fibrosis stage, whereas including the string ratios (F14-19) in the generalized linear models helped to improve late fibrosis stage distinction. The first part of these results is in agreement with the results of a previous study of early stage collagen remodeling 26 , which revealed alterations in collagen shapes but no substantial change in the total amount of collagen. Moreover, the second part of our findings suggests that although the total amount of collagen is altered during late liver fibrosis stages, the ratio of the different collagen string features is a better method for assessing fibrosis stages.

ISHAK fibrosis stages
In this study, the main reason for liver biopsy of chronic hepatitis B patients was pre-treatment evaluation. Therefore, these patients were mostly in the immune clearance stage and were not evenly distributed over all clinical stages. As a result, many patients were positive for HBeAg but were already progressing to liver cirrhosis, suggesting a failure of HBeAg seroconversion despite repeated hepatitis flares. However, even during the immune clearance phase, patients who were positive for HBeAg still had a borderline lower fibrosis score (mean ± SD, 3.42 ± 1.41 vs. 3.06 ± 1.36; HBeAg negative vs. positive, P = 0.156).
In conclusion, this study developed a fully automated and quantitative assessment system employing SHG/TREF microscopy to specifically assess HBV-related fibrosis. We systematically optimized the parameters for quantitative SHG/TPEF liver tissue imaging and developed fully automated image analysis algorithms. Compared with the traditional staining systems, the present system enables enhanced collagen fiber detection and quantification for fibrosis research and clinical diagnosis. SHG/TPEF microscopy is a standard analytical tool, and its use in combination with a computer-based detection and calculation system represents a powerful technique that can be applied in multi-center comparisons of liver fibrosis.

Methods
Tissue Preparation. The study cohort consisted of 175 liver biopsy samples from chronic HBVinfected patients who underwent liver biopsies at LinKo Chang Gung Memorial Hospital between 2000 and 2012. Clinical and pathological characteristics were obtained from the patients' medical records. As described below, liver fibrosis was staged by pathologists according to the ISHAK scoring system and was then automatically evaluated with computational algorithms according to 19 morphological variables. Of the 175 samples, 105 (training cohort) were used for developing a computer-based algorithm with a close correlation to the ISHAK fibrosis score. The other 70 samples were used to validate the scoring method derived from the training system. This study was approved by the Ethics Committee of Chang Gung Memorial Hospital. Written informed consent was obtained from each patient, and the study was carried out in accordance with the approved guidelines.

Masson's Trichrome Staining. Masson's trichrome staining was performed at the Chang Gung
Memorial Hospital Department of Anatomic Pathology as follows. First, 5-micron FFPE sections were deparaffinized and hydrated in distilled water. Then, Bouin's fixative was used as a mordant for 1 h at 56 °C. The FFPE sections were cooled and washed in running water until the yellow coloring disappeared. The samples were stained in Weigert's hematoxylin stain for 10 min, thoroughly washed in tap water for 10 min, stained again in an acid fuchsin solution for 15 min, and then rinsed in distilled water for 3 min. After rinsing, the slides were treated with phosphomolybdic acid solution for 10 min and then rinsed in distilled water for 10 min. Finally, the slides were stained with a light-green solution for 2 min and rinsed in distilled water. After thorough dehydration using alcohol, the slides were mounted, and coverslips were placed onto them. Image Acquisition System. Images were acquired using a Genesis TM (HistoIndex Pte. Ltd, Singapore) system, in which SHG microscopy was used to visualize collagen, and the other cell structures were visualized with TPEF microscopy. Details of the optical settings used can be obtained from Sun et al. 26 and Tai et al. 27 .
The samples were laser-excited at 780 nm, SHG signals were recorded at 390 nm, and TPEF signals were recorded at 550 nm. Images were acquired at 20× magnification with 512 × 512 pixel resolution, and each image had a dimension of 200 × 200 μ m. Multiple adjacent images were captured to encompass large areas. Nine images (3 × 3 images, 600 × 600 μ m dimension) were acquired per site. Overall, five 3 × 3 images from five sites were obtained for each sample. The coefficients β 0 , β 1 , …β n were estimated with the maximum likelihood method. The second method used was the SVM model. A hyperplane was constructed for optimally classifying the subjects into fibrosis stages based on the feature values in the m-dimensional space (m ≥ n). The radial basis function (RBF) kernel was used as the distance measure between two subject vectors x and x', i.e., where σ controls the width of the kernel.