Spatially resolved classification of kidney topography and multicellular composition with deep learning

Introduction: Chronic Kidney Disease (CKD) is a prevalent kidney condition where the kidney gradually loses its ability to filter waste products and surplus fluid effectively. It results in a buildup of toxins in the bloodstream and contributes to other health complications including cardiovascular diseases. Acute Kidney Injury (AKI) is a sudden loss of kidney function. Kidney disease ranks among the prominent causes of mortality in the United States, with approximately 37 million adults suffering from it. Definitive diagnosis of acute kidney injury or chronic kidney disease via histopathological examinations is a labor-intensive process. This study aims to provide a unified platform to assess kidney infections utilizing data from standard histopathological techniques such as trichrome, periodic acid-Schiff, and hematoxylin and eosin staining, and complex techniques such as multiplexed immunofluorescence (mIF) imaging and imaging mass cytometry (IMC).

Materials and

Methods: Extensive studies have been conducted to understand fibrosis and the immune-cell landscape in various kidney diseases. Studies revealed that the trichrome stain explains fibrosis much more effectively than other histochemical (HC) stains. Images acquired through sophisticated techniques such as mIF and IMC yield deeper insights into the distribution of immune cells within their microenvironment and structural characteristics within biopsies, that traditional methods cannot provide. The proposed pipeline combines the advantages of both techniques to understand the disease.

The study utilizes the dataset created by the Kidney Precision Medicine Project, particularly the images acquired through HC staining and mIF imaging. The proposed methodology is split into sections based on the input type. Graph-based approaches have been developed to study these images.

The mIF pipeline is designed to extract cells (MESMER instance segmentation) and construct a biomarker expression matrix along with spatial information; this matrix is used to perform cell phenotyping. Structures of the nephron are then extracted (semantic segmentation) from the corresponding biomarkers and will be utilized to study the immune-cell landscape around different topological structures of the kidney. A graph-based binary classifier was employed to understand this landscape around the topological structures, and its performance was benchmarked against classical machine learning and neural network models.

Similarly, an extra-cellular-matrix-based graph is constructed from trichrome whole slide images (extraction of topological structures and interstitial fibrosis regions) in the trichrome HC pipeline. These graphs were used to categorize the patients, and the graph-based binary classifier’s performance is benchmarked similarly.

Results, Conclusions, and Discussions: The proposed architecture for binary classification of patients into AKI or CKD produced an accuracy of (70.63 ± 3.13) % and an AUC score of 0.8759 ± 0.0202 when the input to the neural network is a graph of immune cells present in the entire tissue sample/patch. The model was trained on 94 AKI patches and 141 CKD patches. Due to the class imbalance, AUC will be the metric used to compare models’ performances. The model outperformed the other models, showing the highest AUC score, second highest accuracy (after random forest), and the least variances in accuracy and AUC score. When the model’s input was the structure-interstitial fibrosis (IF) region graph, the model reported an accuracy (51.11 ± 15.96) % and an AUC score of 0.6667 ± 0.1581. The model is the second best in terms of mean accuracy and AUC score (after Support Vector Machine), and has the least variance in AUC scores when compared to other models. The training dataset contained 20 CKD graphs and 21 AKI graphs.

The models used here provide insight into the distributions and densities of various cells or structures whose count or areas explain the progression of kidney disease. While all types of cells served importance during classification, the mIF binary classification model stated that the density of macrophages, monocytes, and podocytes in the two populations differ (most notable features). The trichrome binary classification model states that the density of tubules, IF regions, and sclerosed glomeruli helped distinguish the two populations.

The multimodal novel study incorporates information from diverse sources that provide valuable information on different cells or parts of the kidney, which is better than a sole source. Future work includes utilizing these models that provide insight into heterogeneous graphs that combine immune-cell information along with interstitial fibrosis regions (current graphs are homogenous, i.e., either mIF-cells-based or trichrome-region-based) and provide a deeper understanding of the immune-cell landscape in areas of fibrosis. This can be a potential direction for understanding the progression of diseases in kidneys after their onset.

Acknowledgements (Optional):