An automatic deep learning platform for kidney transmission electron microscopy image analysis

About This Grant

Summary 37 million people in the US suffer from chronic kidney disease (CKD). Unfortunately, the multifaceted burden of CKD (prevalence, morbidity, mortality, costs) is relentlessly growing. CKD progression and proteinuria, a biomarker of this process, are associated with injury to the kidney's glomerular filtration barrier (GFB), including the glomerular basement membrane (GBM) and podocyte foot processes (PFPs). These nanometer-scale structures can only be resolved visually using transmission electron microscopy (TEM). TEM images of kidney tissues are commonly used to detect morphological changes in the GFB associated with disease. In both research and clinical settings, GBM and PFP width measurements taken from TEM images of kidney tissue are used as image-based biomarkers - aiding the diagnosis and classification of glomerular diseases. Currently, these imaging biomarker measurements in TEM images are performed manually, which is labor-intensive and introduces significant inter-operator variation and bias, limiting glomerular disease research and clinical diagnosis. At present, no tools can achieve accurate results close to manual measurements. Significant variance across different diseases, the lack of high-quality labeled training data, high-resolution images with multi-scale and fine-grained features, and boundary ambiguity make TEM image analysis extremely challenging for general AI models. Thus, a specially designed narrow AI model is needed to efficiently and accurately solve these challenges. To fill this emerging gap, we propose a novel deep learning framework integrating self-learning and a combined High-Resolution and Pyramid Pooling Module architecture, specifically designed to address these challenges. Images with different magnifications generated from various institutes will be utilized to improve the model's generalization. An independent set of manually labeled images from humans, mice, and rats will be used to evaluate segmentation accuracy, particularly for GBM and PFPs. Considering that most biomedical researchers and physician scientists in the kidney clinical and research community lack sufficient computational skills or resources to execute deep learning models, our system will be designed with an intuitive and user- friendly interface tailored to meet the needs of non-computational users. The web-based interface will be available on publicly accessible cloud platforms for users without proper computational resources, and the source code and instructions will be released on a GitHub repository to enable users to install the system on their institute's internal server. In addition, with the support of pathological groups from several different institutes, we will apply the proposed method and tool to more than 70,000 TEM images to further validate its efficacy and generalizability.