2.1. Patient Recruitment
We retrieved 20 patients diagnosed with membranous nephropathy by clinical data and renal biopsy in the China-Japan Friendship Hospital from July 2019 and September 2019, including 10 IMN patients and 10 HBV-MN patients. The inclusion criteria of the IMN group were MN patients with unclear etiology and glomerular lesions limited to the immune complex deposited under the epithelial and thickening glomerular basement membrane. The HBV-MN group had the following criteria: (1) serum HBV markers positive; (2) excluded other causes attributed to secondary renal disease (SLE, drugs, toxins, other infections, or malignancy); and (3) presence of detectable HBV-related antigen or antibody in renal biopsy tissue. This last criterion is the most fundamental and indispensable rule of all of those listed above. In all of the cases, MN was accompanied by other pathological patterns, and diabetic nephropathy and IgA nephropathy were ruled out. Demographic and clinical parameters of the MN patients are shown in Table 1.
Table 1. Demographic and clinical parameters of the 20 patients
Characteristic
|
IMN (n=10)
|
HBV-MN (n=10)
|
Age (years)
|
47.6±14.6
|
50.3±11.0
|
Men, n (%)
|
6 (60%)
|
7 (70%)
|
Proteinuria (g/24 h)
|
4.88(2.00, 6.95)
|
3.74(3.03, 5.36)
|
Albumin (g/L)
|
30(25, 39)
|
30(25, 31)
|
Serum creatine (μmol/l)
|
66.5(52.2, 83.6)
|
75.5(55.2, 98.7)
|
BUN (mmol/l)
|
4.67(3.34, 5.66)
|
4.85(3.73, 6.63)
|
Cholesterol (mg/dl)
|
6.34(5.13, 8.50)
|
7.46(6.23, 8.94)
|
PLA2R-ab (Ru/ml)
|
29.9(18.0, 56.8)
|
7.6(4.8, 15.9)
|
Abbreviations: IMN idiopathic membranous nephropathy, HBV-MN hepatitis B virus-related membranous nephropathy, BUN blood urea nitrogen, PLA2R M-type phospholipase A2 receptor.
2.2. Sample Description
We collected glomerular characteristics for hyperspectral analysis with routinely processed renal biopsy tissue. Each sample was stained with hematoxylin-eosin (HE), periodic acid-Schiff (PAS), Masson's trichrome, and Jones's silver. All renal tissue samples were observed by light and immunofluorescence microscopy beforehand, and all of the patients' diagnosis was confirmed based on current criteria. Later an experienced expert re-examined these biopsies, and all of the patients were eligible for consideration.
2.3. Hyperspectral Image Collection
We performed hyperspectral imaging using a compound microscope system, where a pushbroom (line-scanning) imager SOC-710VP (Surface Optics Inc.) is employed combining with a microscope (Olympus CX31RTSF). The imaging system captures spatial size 696x520 with 128 spectral bands, covering the spectral range from 400nm to 1000nm with a spectral resolution of 4.69 nm. The heart of the HS imager consists of a spectral dispersion element and a 2-dimensional focal plane array (FPA) detector. In our system, the dispersive spectrometer is a diffraction grating where the incoming light produced from the microscope is separated into discrete wavelengths before being projected onto the detector. Next, the charge-coupled device (CCD) detector is activated to capture the intensity of the light at each pixel of the image using the HyperSpectTM operating software. For each patient, we randomly selected 2–3 glomeruli per slide, and then manually marked out every immune complex in the subepithelial area using the ENVI 14.0 software before exporting the data for further analysis.
2.4. Image De-noising
Unprocessed HSI data usually contains high spectral noise generated by the imaging system. This noise can lead to undesirable effects. In order to remove the noise in the data, a mean filter was applied. Figure 2 shows the concept of mean filtering is to replace the center pixel with the average value of all of the pixels inside the local window, reducing the amount of intensity variation between one pixel and its neighbors. Here, the de-noising process is achieved via the following equation (1). Where D(i, j) is the value of the center pixel in the filtering window, S(m, n) is the value of pixels in the window (i=0,1…H-1; j=0,1…W-1), and W and H represent the width and height of the filtering window, respectively.
Figure 3 shows a glomerulus from the 10th channel of an HBV-MN patients' HS image before and after de-noising. The remarkable improvement confirms the effect of reducing the system noise of HSI data.
2.5. Projection Transformation
HSI data has hundreds of spectral information channels; hence, an essential step for utilizing HSI data is to reduce the redundant information in its spectral signature. Projection transformation is an advanced method for acquiring the maximum reduced subspace of a target without losing its essential information. Current projection transformation techniques include principal component analysis (PCA), independent component analysis (ICA), and Fisher's linear discriminant analysis (LDA). However, a significant drawback of those techniques is that they all are only legitimate when the target data has a Gaussian distribution. In this study, we developed an alternative method named local Fisher's discriminant analysis (LFDA). The typical LFDA projection is calculated by maximizing Fisher's ratio and using these local scatter matrices. The biggest benefit of LFDA is that it can obtain good between-class separation in the feature subspace while preserving the within-class local structure [16]. It also integrates the advantages of both Fisher's linear discriminant analysis (LDA) and locality-preserving projections (LPP) while bypassing the requirement for a Gaussian distribution [17-20]. Figure 4 (a) and (b) are the before and after projection transformation feature distribution of one testing sample. The results exhibit the effectiveness of LFDA for seeking a subspace with maximum separability for features.
2.6. Proposed Deep Learning Framework
Following the image de-noising and projection transformation procedures mentioned above, we constructed a deep neural network (DNN) to extract and classify the intrinsic and high-level features of the different glomerular images.[21] More specifically, support vector machine (SVM), extreme learning machine (ELM) [22], Alexnet [23], Resnet20[24], and VGG19[25] were implied on the MN database with and without the pre-processing procedures to achieve the ultimate goal of formulating an MN identification architecture that can automatically distinguish HBV-MN from IMN. For the deep learning models: First, Leaky ReLU activation function is applied after each convolutional layer to solve the problem of gradient vanishing and accelerate the fitting speed; second, batch normalization (BN) strategy behind some convolutional layers is used as a regularizer to simplify the tuning process and lower initialization requirement; third, dropout technique is employed to avoid the over-fitting issue, and the dropout rate is set to 0.5. The proposed DL models were designed and developed using PyTorch.
The validation of the proposed DL framework was performed using leave-one-out cross-validation (LOOCV), where samples of the patients to be tested were extracted from the database before training the algorithm. This methodology guarantees the database for the training and testing process are strictly separated; it also proves the eligibility and consistency of comparing the performance between each method.
Here, we applied DNN to identify glomerular disease in microscopic hyperspectral images for the first time, and then verified and supplemented the outcome of immunofluorescence or light microscopy.