Periodic Reporting for period 2 - AIM.imaging.CKD (AI-augmented, Multiscale Image-based Diagnostics of Chronic Kidney Disease)
Okres sprawozdawczy: 2022-11-01 do 2024-04-30
The overall goal of AIM.imaging.CKD is to specifically address this unmet need by developing, validating, and integrating image-based diagnostics for CKD, which could ultimately lead to improved kidney disease patients’ management and outcomes.
Specifically, we aim to develop a multiscale approach from nano- to micro- to macromorphological and molecular diagnostics. This will include augmented full-spectrum ultrastructural (“nano”) and histological (“micro”) renal biopsy diagnostics, focusing on reproducible, quantitative analyses and prediction of clinically relevant outcome parameters. We will also explore macro-morphological and molecular imaging in CKD, focusing on translatable non-invasive approaches. For all these technologies, which produce diagnostic medical images, we leverage advanced image analysis models, particularly using deep learning technologies.
One of the highlights we have achieved so far is the development of Next-Generation Morphometry for pathomics-data mining in kidney histopathology. Here, we proposed a deep-learning (DL) based framework that automatically and reproducibly extracts large-scale digital, quantitative, and interpretable morphological biomarkers from patients’ histopathology in a high-throughput manner and is even applicable in rare diseases. This represents a novel methodological approach for unbiased large-scale data mining of histomorphology, which we term “Next Generation Morphometry - NGM”, providing a novel “omics” approach, which we term “pathomics”. Similarly to “next-generation sequencing - NGS” and “genomics”, which previously revolutionized molecular pathology diagnostics, our proposed framework for next-generation Morphometry (“NGM”) might have a similar potential to advance morphology-based pathology diagnostics. Using this method enabled us to comprehensively analyze the fates of kidney structures during recovery from kidney injury, revealing that the current available kidney functional parameters do not well reflect the ongoing pathological processes and limited regenerative capacity of the kidney. This substantiates the need to develop non-invasive diagnostic imaging approaches, one of the focuses of this project.
Another example is the area of assessment of kidney transplant biopsies, which is currently the only means to diagnose allograft diseases and particularly transplant rejection. The correct diagnostics of kidney transplant biopsies guide the treatment and management of patients. At the same time, transplant biopsy diagnostics is one of the most complex and time-consuming areas of pathology and is still plagued by considerable variability, some of which are therapeutically relevant. We here addressed this unmet need, and to our knowledge for the first time, developed, tested, and validated deep-learning support systems to augment the histopathological diagnostics of kidney transplant biopsies. This was the largest transplant deep learning study in pathology, bringing together a strong international and interdisciplinary team, focusing on relevant clinical application scenarios. This study opened a new research area and strengthened our position as an internationally leading team in deep learning development in kidney pathology.
The amount of data volume is steadily increasing in medicine (including pathology), with the largest annual growth between all major sectors. Deep learning requires considerable computational resources, and newer models are increasingly deep. This is particularly true for large language models such as ChatGPT. While there is consensus on the potential role and value of deep learning in medicine, the potential ecological consequences have not yet been considered. We have performed the first sustainability study on the ecological consequences of the implementation of DL in pathology, showing potential severe impact (within our center and extrapolated to national international, and future scenarios). Given that this could limit the use of deep learning, we also explored and proposed potential solutions for green and sustainable AI. This study might serve as a blueprint to consider sustainability in development and implementation in pathology, and potentially in medicine in general.
The project started early during the COVID-19 pandemic. While this was not the original focus, we were able to support COVID research and show the potential implications of approaches and techniques developed in this project beyond kidney diseases. Facilitated by the National Autopsy Network that we initiated and are coordinating (https://naton.network/) we have revealed potential mechanisms of how SARS-CoV-2 drives lung fibrosis.
From an infrastructural perspective, with institutional support and in part driven by this project, we were able to secure finances for a module building solely dedicated to digital (nephro)pathology. This building will encompass and harbor the whole digital pathology workflow, strongly facilitating this project and future developments of our group in this area.
In an outlook, we expect to achieve our goals by the end of the project.