​Chronic Kidney Disease (CKD): Are You Suitable For Single And Dual Kidney Transplants?

for more information:ali.ma@wecistanche.com

Part Ⅰ：A neural network for glomerulus classification based on histological images of kidney biopsy

Giacomo Donato Cascaranol,Francesco Saverio Debitontol & et al.

Background

Chronic Kidney Disease (CKD) is a pathological condition characterized by a functional degeneration of the kidney. CKD (Chronic Kidney Disease) is the 12th cause of death, with up to 1.1 million cases worldwide; the increased mortality related to CKD (Chronic Kidney Disease) of the last years makes it one of the fastest rising causes of death, alongside diabetes and dementia [1, 2]. Kidney transplantation is the best renal replacement therapy as revealed to be more effective than dialysis treatment in terms of long-term mortality risk and, at the same time, has a reduced impact on the public health system [3, 4].

Click to Cistanche NZ for Chronic kidney disease

Liyanage et al. estimated that 2.6 million people, in the face of 4.9 million patients, received renal replacement therapy worldwide in 2010, suggesting that at least 2.3 million people might have died prematurely because appropriate therapy could not be accessed [5].

Due to the increasing necessity of kidney transplants [6], different studies tried to widen the criteria for accepting kidneys for being transplanted, which are generally excluded based on the donor's age and other characteristics related both to the quality and dimension of kidneys [7,8].

Moore et al. performed a comparison between dual kidney transplantation from Expanded Criteria Donors (ECDs) and single kidney transplantation from concur-rent ECDs and standard criteria donors. The authors assessed that the use of dual kidney transplantation from marginal donors is a viable option and that renal function can be achieved, provided that both kidneys are transplanted into a single recipient [9].

Remuzzi et al. proposed a technique to assess the kidney condition by evaluating histological biopsies [10]. The evaluation criterion, known as the Karpinski score, considers the evolution (in percentage)of a pathological condition of four main functional areas: glomerulosclerosis, tubular atrophy, interstitial fibrosis, and arterial sclerosis. This score ranges from 0 to 12, and the higher the number, the worse is the kidneys condition [10-12]. Kidneys with a Karpinski score from 0 to 3 and from 4 to 6 are considered suitable for single and dual transplants, respectively.

Treatment of kidney disease: cistanche and kidney transplantation

To assess the Karpinski score, pathologists perform the visual evaluation of the histopathological Whole-Slide Images(WSIs). This process is usually time-consuming, prone to error, and also subjective.

To overcome these drawbacks, the development of Computer-Aided Diagnosis(CAD) systems based on histopathological tissue image analysis for supporting the computation of the score is a valuable headway.

Recent literature works show the application of image processing and machine learning techniques to analyze kidney histopathological WSIs for glomeruli detection and classification. Image processing approaches aim to extract meaningful features, e.g., those based on shape and texture analysis; then, machine learning algorithms, such as shallow or deep Artificial Neural Networks (ANNs), make decisions based on extracted features.

Simon et al, for example, proposed texture-based features set as a simple but effective automatic method for glomeruli localization [13]. The authors applied the algorithm on renal tissue sections and biopsies of large histopathological WSIs. The features extracted from an adaptation of the Local Binary Pattern (LBP) algorithm were used to train a Support Vector Machine(SVM)model. The authors reported high precision(>90%) and reasonable recall (>70%) as results.

To perform a comprehensive detection of glomeruli in images of whole kidney sections, Kato et al. proposed a new descriptor called Segmental HOG (Histogram of Oriented Gradients)[14]. The authors claimed the robustness of the solution and high-quality segmentation outputs; furthermore, the authors compared Segmental HOG with Rectangular HOG showing that the first approach reached significant improvements in detection performance.

Several authors, instead, focused on the analysis of glomeruli's shape and color. Kotyk et al.proposed a novel solution to face the wide intensity variation and the inconsistency in terms of shape and size of the glomeruli in the renal corpuscle. The proposed approach, based on the Particles Analyzer technique, allowed the detection of the renal corpuscle and the following measurement of glomerulus diameter and Bowman's space width. The authors assess that the approach was robust to glomeruli deformations even with glomerular hypertrophy [15]. An analysis of the effects of significant diversity of color and tissue shape on whole slide images was performed by Zhao et al. [16]. The authors focused on the extraction of Bowman's capsule width to design an automated glomerulus extraction framework from the micrograph of the entire renal tissue. The system was tested on non-human primates renal tissues with Haematoxylin and Eosin(HE)staining.

Bukowy et al. proposed a different analysis workflow. In [17], the authors developed a convolutional neural network to detect glomeruli in trichrome-stained kid-ney sections. The procedure was tested on rat kidneys and the reported results, regarding the classification of healthy and damaged glomeruli, show average precision and recall of 96.94% and 96.79%, respectively.

In a previous work by Bevilacqua et al, a CAD system for segmentation and discrimination of blood vessels ver-sus tubules from biopsies in the kidney tissue has been designed and tested [18]. Histological images with Peri-odic Acid-Schiff(PAS)staining have been used to segment Regions of Interest (ROIs) and extract Haralick features allowing a subsequent classification procedure using algorithms based on ANNs. Test results determined that the supervised ANN approach was consistent, allowing obtaining good classification performance.

This work focuses on the automatic evaluation of kidney biopsies, dealing with a specific pathological condition considered by the Karpinski score: glomerulosclerosis, ie. the ratio between sclerosed glomeruli and the overall number of glomeruli. To do this, the detection and discrimination of the sclerotic condition affecting the glomeruli from those non-sclerotic are crucial. As already reported in works from the state-of-the-art, this is a challenging task due to the glomeruli wide intensity variations and inconsistencies in shape and size.

treatment of kidney disease: single and dual transplants researches

A combination of different feature extraction algorithms has been designed and evaluated for discriminating the condition of glomeruli. The reported literature shows specific and unique image processing algorithms applied on different types of staining and non-human WSIs. The set of features proposed in this work, instead, comes from a collection of two wide-used, well-known, and general-purpose features extractor algorithms families,i.e.morphological and texture features. These feature families are also included in some of the algorithms proposed in the literature, but in this work, they were extracted from human WSIs with PAS staining. In addition, the classification pipeline, detailed in Methods, includes also procedures for features reduction allowing the design of a shallow Artificial Neural Network. The overall workflow proposed in this work, and the integration with the procedure presented in[18], will allow us to build up a complete CAD system for the analysis of histopathological WSIs.

Results

The results obtained by evaluating the proposed classification workflow on the test set are reported. In particular, results refer to the performance obtained considering the reduced set of features classified by using the cross-validated shallow ANN. As reported in Table 1, the test set was constituted by 579 glomeruli images: 87 sclerotic, 492 non-sclerotics.

To evaluate the workflow stability, 10 runs of the entire process were performed. The achieved results are summarized in Table 2. In particular, the results are reported in terms of mean and standard deviation of several metrics,i.e.Accuracy (Eq.1), Precision(Eq.2), Recall (Eq.3), and Matthews Correlation Coefficient (Eq.4)[19], evaluated according to the confusion matrix reported in Table 3.

Among the iterations, the best results are reported in Table 4, whereas the corresponding confusion matrix is reported in Table 5.

The implemented workflow allows the classification of sclerotic and non-sclerotic glomeruli with good performances(mean MCC=0.95 and mean Accuracy=0.99)and low variability (MCC std=0.01 and Accuracy std<0.00)(see Table 2). Precision and Recall are equal to 0.98 and 0.93, respectively, thus showing that the proposed system achieves a better performance in the non-sclerotic evaluation(all the non-sclerotic glomeruli were detected in the best case).

Table 1 Dataset configuration	Table 2 Metrics comparison of 10 network initializations

Table 3 Confusion Matrix for metrics computation	Table 4 Metrics comparison of 10 network initialization

Table 5 Confusion matrix of the best model

Discussion

Evaluating the proposed approach on an independent test set, the classification workflow achieved a mean MCC and Accuracy of 0.95 and 0.99, respectively, and low variability over 10 independent iterations (MCC std=0.01 and Accuracy std<0.00). Good precision and recall were also obtained (Precision: 0.9844±0.0111, Recall: 0.9310±0.0153). The proposed approach thus leads to an improvement of the classification performance if compared to the reported literature[13, 17].

While implementing and evaluating the reported workflow, we faced and tested the common data unbalancing problem, which has been solved by using MCC as performance comparison coefficient and ROC curve for selecting the optimal classification threshold. The reported results suggest that the proposed workflow setup is reliable for the investigated domain, supporting the clinical practice of discriminating the two classes of glomeruli.

Analyzing misclassified glomeruli, we found also that the input images corresponding to the misclassified samples showed staining artifacts or partial parts (mostly on the edges); common examples are mentioned in Fig.1.In the clinical practice, however, pathologists discard such images which could also be excluded in the proposed workflow by designing strategies for detecting in advance images affected by such problems.

Treatment of kidney disease: single and dual kidney transplants

CLICK HERE TO PART Ⅱ