2.1.

Tissue Samples

The ex vivo matured human skin scar specimens (30 normal and 30 abnormal scars) were collected during plastic surgery performed on 60 Chinese female patients. The scars were diagnosed based on histological examination by certified pathologists and classified according to the traditional histopathological criteria.^4,18 Prior to study participation, all patients signed an informed consent, and this study was approved by the Institutional Review Board of Fujian Medical University. To protect the tissue samples from metamorphism, we stored them in liquid nitrogen ($-196°\mathrm{C}$) immediately after they were excised from patients. Prior to SHG imaging, the tissues were sectioned in $140\text{-}\mu \mathrm{m}$ thicknesses and sandwiched between a microscope slide and cover glass. To avoid dehydration or shrinkage during the imaging acquisition, a little phosphate-buffered saline (PBS) solution was dripped into the tissue specimen.

2.2.

Second-Harmonic Generation Microscopy

The SHG imaging was performed on a Zeiss LSM 510 META laser scanning microscope coupled to a mode-locked femtosecond Ti: sapphire laser (Coherent Mira 900-F) operating at 810 nm. The sapphire laser (110 fs, 76 MHz) was used as the excitation light source. An oil immersion objective ($\times 63$ and $\mathrm{NA}=1.4$) was employed for focusing the excitation beam into tissue samples (average power less than 5 mW) and for collecting the backscattered intrinsic SHG signals. The high-resolution ($512\times 512\text{pixels}$) image was obtained at $2.56\mu \mathrm{s}$ per pixel (total: 1.57 s) with a view field of $230.8\mu \mathrm{m}\times 230.8\mu \mathrm{m}$. In this study, one SHG image was collected from each patient, resulting in a total of 60 scar collagen SHG images (30 normal and 30 abnormal) for analysis.

2.3.

Texture Analysis

2.3.1.

Homogeneity features extracted by local difference local binary pattern

To characterize the homogeneity features in an image, the local contrast information derived from LD-LBP was used. Given a gray value of central pixel $g_c$ and its neighbors $g_p$ ($p=\mathrm{0,1},\dots ,P–1$, $P$ is the number of sampling points in a neighborhood), the difference between $g_c$ and $g_p$ can be calculated by $C_p=g_p-g_c$, and the amplitude of $C_p$ is expressed as $S_p=|C_p|$. As an example, Fig. 1(a) shows the $3\times 3$ neighborhood with central pixel $g_c$ being 20 and the neighboring pixels $g_p$ are [12, 71, 43, 39, 27, 31, 14, and 6]. The differences in $C_p$ [Fig. 1(b)] are [-8, 51, 23, 19, 7, 11, $-6$, $-14$], and the amplitudes $S_p$ are [8, 51, 23, 19, 7, 11, 6, 14], as shown in Fig. 1(c). Then LD-LBP is defined as

Eq. (1)

$$\mathrm{LD}\text{-}\mathrm{LBP}=\sum _{p=0}^{P-1}H(S_p-t)2^p,$$

where $H(x)=\{{\displaystyle \genfrac{}{}{0ex}{}{1,}{0,}}{\displaystyle \genfrac{}{}{0ex}{}{x\ge 0}{x<0}}$ and $t$ is a threshold to be adaptively determined. Here, we set $t$ as the mean value of $S_p$ from the whole image. The LD-LBP operator is defined in a consistent format with that of the basic LBP operator. Similar to the basic LBP, the rotation invariant version of LD-LBP is also defined to achieve rotation invariant classification. Additionally, the LD-LBP uses a rotation invariant variance measure VAR that characterizes the contrast of local image texture as the representation of homogeneity features

Eq. (2)

$$\mathrm{VAR}=\sum _{p=0}^{P-1}{(g_p-w)}^2/P,$$

where

Eq. (3)

$$w=\left(\sum _{p=0}^{P-1}{g}_p\right)/P.$$

Fig. 1

(a) $3\times 3$ neighborhood with central pixel being 20; (b) the local differences; and (c) the amplitudes of local differences.

Low variance (VAR) occurs in images with homogeneous features, otherwise the VAR is higher.

2.4.

Classification

An SVM classifier was used to test the ability of the texture parameters to classify the SHG images into normal or abnormal.^23–25 Given a training sample set ${[(x_i,y_i)]}_{i=1}^n$, where $x_i$ denotes the training vector, $x_i\in R^n$ and $y_i$ denote the corresponding class label, $y_i\in \{1,-1\}$, and $n$ denotes the total number of the training sample. SVM finds the solution of the following optimization problem

Moreover, the LOOCV procedure was used to address the drawback of the small-sized sample.^26,27 Each time one single sample was considered as the validation data and the other samples were considered as the training data. In this work, we have a total of 60 scar collagen SHG images (30 normal and 30 abnormal). The procedure was repeated 60 times until every sample was used as the validation data.

2.5.

Statistical Analysis

For statistical analysis, a total of 30 normal and 30 abnormal scar collagen SHG images were measured for texture analysis. Statistical analyses were performed by using the MATLAB® (version 2010b) software. Quantitative data were presented as a mean value with its standard deviation indicated ($\text{mean}\pm \mathrm{SD}$) and compared using the Student’s t-test analysis. Differences were considered to be statistically significant when the $p$-values were less than 0.05.

3.1.

Collagen Second-Harmonic Generation Images of Human Skin Scars

Representative high-resolution collagen SHG images from human normal and abnormal scars are shown in Figs. 2(a) and 2(b), respectively. This presents a great difference in the morphological structure of the collagen between the normal and abnormal scars. As we can see from Fig. 2(a), large and coarse collagen bundles with a well-regulated distribution and well-defined orientation in the normal scar are well identified. In contrast, a less organized and disordered collagen structure characterized by thinner but more complex fibril bundles are clearly seen in the abnormal scar [Fig. 2(b)]. All these observations are well consistent with the previous description of normal or abnormal scars.^28,29

Fig. 2

Collagen second-harmonic generation (SHG) images in human (a) normal and (b) abnormal scars. Images are $512\times 512\text{pixels}$.

3.2.

Quantification of Collagen in Human Skin Scars

In order to quantitatively characterize the collagen morphology in normal and abnormal scars such as collagen distribution, collagen orientation, and collagen coarseness, we applied detailed texture analysis based on LD-LBP and wavelet transform to collagen SHG images. Texture parameters including homogeneity, directional and coarse features of SHG images were extracted. Demonstrated in Fig. 3 is the decomposition process of the collagen SHG images, including an original collagen SHG image of normal scar [Fig. 3(a)], the corresponding LD-LBP transformed image in gray scale [Fig. 3(b)], and the four subimages by one-level wavelet transform decomposed from the LD-LBP image [Fig. 3(c)].

Fig. 3

The decomposition process of collagen in an SHG image. (a) Original collagen SHG image of normal scar. (b) The local difference local binary pattern (LD-LBP) transformed image in gray scale from (a). (c) The four subimages by one-level wavelet transform decomposed from (b).

3.2.1.

Homogeneity feature

Figure 4 presents the LD-LBP variances of collagen SHG images from normal and abnormal scars, showing a much smaller value of normal scar SHG images compared to abnormal scar SHG images ($0.04\pm 0.01$ versus $0.55\pm 0.35$; $p<0.05$). Quantitatively analyzed result indicates that the normal scars have a more homogeneous collagen morphology than the abnormal scars.

Fig. 4

LD-LBP variances of collagen SHG images from normal and abnormal scars. Error bars indicate calculated standard deviations.

3.2.2.

Directional and coarse features

Figure 5 gives the energy values in each subimage on both normal and abnormal scar SHG images. By comparing energy values in each subimage, it can be seen that the energy is mainly concentrated in the LL subimage compared to LH, HL, or HH subimage, for both normal ($2.16\pm 0.09$ versus $0.44\pm 0.03$, $0.26\pm 0.04$, or $0.20\pm 0.03$; $p<0.05$) and abnormal scar collagen SHG images ($1.72\pm 0.11$ versus $0.69\pm 0.09$, $0.32\pm 0.02$, or $0.18\pm 0.02$; $p<0.05$). Moreover, texture details are mainly distributed in the LH subimage and less distributed in HL or HH subimages for both normal ($0.44\pm 0.03$ versus $0.26\pm 0.04$, or $0.20\pm 0.03$; $p<0.05$) and abnormal scar collagen SHG images ($0.69\pm 0.09$ versus $0.32\pm 0.02$, or $0.18\pm 0.02$; $p<0.05$).

Fig. 5

Energy values in each subimage of collagen SHG images from normal and abnormal scars. Error bars indicate calculated standard deviations.

Energy ratios in the low-frequency subimage are depicted in Fig. 6. This clearly shows that energy ratios in the low-frequency subimage of normal scar SHG images are higher than those of abnormal scar SHG images ($70.55\pm 1.93\%$ versus $59.27\pm 0.46\%$; $p<0.05$). These statistical results are consistent with the observation that the collagen arrangement within normal scars is ordered and coarser compared to abnormal scars.

Fig. 6

Energy ratios in low-frequency subimage of collagen SHG images from normal and abnormal scars. Error bars indicate calculated standard deviations.

3.3.

Optimal Wavelet and Decomposition Level

Table 1 describes the accuracy of the SVM classifier in the LOOCV procedure based on texture parameters of LD-LBP combined with Haar, Daubechies, Symlets, and Morlet wavelet transform. As can be seen from the table, the overall accuracy of the SVM classifier is relatively higher when LD-LBP is combined with the Haar wavelet. Moreover, Haar wavelet is fast in speed and describes detailed features,³⁰ which were adopted in this study.

Table 1

Accuracy of support vector machine (SVM) classifier in the leave-one-out cross-validation (LOOCV) procedure based on texture parameters of local difference local binary pattern (LD-LBP) combined with Haar, Daubechies, Symlets, and Morlet wavelet transform.

Listed in Table 2 is the accuracy of the SVM classifier in the LOOCV procedure based on the texture parameters of LD-LBP combined with an $n$-level Haar wavelet transform ($n=1$, 2, 3). It can be observed that the SVM classifier has an overall accuracy of 90.0% when LD-LBP is combined with one-level Haar wavelet. The accuracy drops to 73.3% and 71.7% when a two-level or three-level Haar wavelet was used. Therefore, a one-level Haar wavelet transform was selected in this analysis.

Table 2

Accuracy of SVM classifier in the LOOCV procedure based on texture parameters of LD-LBP combined with one-level, two-level, and three-level Haar wavelet transform.

3.4.

Comparison with Other Methods

To evaluate the effectiveness of the proposed method for scar discrimination, the SVM classifier in the LOOCV procedure was trained and applied to classify collagen SHG images of normal and abnormal scar tissues using GLCM, Haar wavelet transform, basic LBP combined with Haar wavelet transform, and the proposed method combining LD-LBP and Haar wavelet transform texture parameters, respectively. Additionally, a multipurpose image classifier called WND-CHARM was also used for comparison.³¹ Results are shown in Table 3. Using the GLCM,^32,33 four texture features including contrast, correlation, energy, and homogeneity were extracted from the scar collagen SHG images and analyzed using the texture average of four directions (0 deg, 45 deg, 90 deg, 135 deg) with distance, $d=10$. It can be seen from Table 3 that the SVM classifier has a higher accuracy when combined LD-LBP and Haar wavelet transform parameters were included in the classification model, while a lower accuracy is obtained when GLCM or Haar wavelet transform was independently used, respectively. The basic LBP combined with a Haar wavelet transform does not give a satisfactory performance for distinguishing abnormal scars from normal scars. In addition, WND-CHARM was less effective than the proposed method specifically designed for texture classification of scar SHG images.

Table 3

Comparison of the accuracy of LD-LBP + Haar wavelet transform to other methods.

Figure 7 compares the receiver operating characteristic (ROC) curves³⁴ with the LOOCV procedure generated by texture parameters from the GLCM, Haar wavelet transform, basic LBP combined with Haar wavelet transform, and the proposed method. The area under the ROC curve (AUC) is a measure of the overall diagnostic performance.³⁵ It is seen from Fig. 7 that the accuracy of the SVM classifier, as represented by the AUC, is determined to be 0.96 when LD-LBP was combined with Haar wavelet transform parameters for classification. These values drop to 0.75, 0.63, and 0.59 when the combined basic LBP and Haar wavelet transform parameters, the GLCM parameters or the Haar wavelet transform parameters were independently used, respectively. The diagonal line $y=x$ is the expected ROC curve of a random guesser (with $\mathrm{AUC}=0.5$).

Fig. 7

Receiver operating characteristic (ROC) curves for texture parameters from the four texture analysis methods including gray level co-occurrence matrix (GLCM), Haar wavelet transform, basic LBP + Haar wavelet transform, and LD-LBP + Haar wavelet transform. The area under the ROC curve is 0.96 for texture parameters (LD-LBP + Haar), 0.75 for texture parameters (basic LBP + Haar), 0.59 for Haar wavelet parameters and 0.63 for GLCM parameters.