Artificial Intelligence-based Image Processing Methods to Advance the Characterization of Polycystic Kidney Disease

Overview

The primary aim of this observational exploratory study is to develop AI-based image processing methods to advance the characterization of Polycystic Kidney Disease using medical images and associated clinical data, including: 1. AI-based fully automatic segmentation techniques for the accurate identification of kidneys, liver, and cysts, with a focus on AI interpretability and robustness; 2. advanced AI-based image processing techniques allowing to identify new imaging biomarkers, including through the use of radiomics, to characterize ADPKD tissue microstructure and therefore stage the disease and monitor and predict disease progression and response to therapy; 3. multiparametric models including image-based radiomic features alongside clinical and laboratory data to stratify ADPKD patients and predict ADPKD progression over time. The study will also have the secondary aim of validating the novel techniques against gold standard (manual) methods, when available.

Autosomal Dominant Polycystic Kidney Disease (ADPKD) is the most prevalent hereditary kidney disease, affecting 12.5 million people worldwide in all ethnic groups. ADPKD is caused by a gene mutation in either PKD1 or PKD2, that leads to the formation and growing of multiple fluid-filled cysts in the kidneys and often the liver, leading to chronic kidney injury and ultimately end-stage renal disease. In ADPKD, kidney function may remain normal for several decades and is therefore not fully informative. The identification of early biomarkers able to accurately monitor and predict disease progression in order to take prompt action and select targeting treatment options is urgently needed. Total kidney volume has been recognized as a prognostic biomarker to select patients for clinical trials and is acknowledged by the scientific community as a relevant biomarker to monitor disease progression and response to therapy, and to predict ADPKD course. Total kidney volume can be quantified using medical images, such as Ultrasound, Computed Tomography, and Magnetic Resonance Imaging (MRI). Renal non-contrast enhanced MRI, denoted by high resolution and with no need for contrast agents or ionizing radiation, is the most suited to monitor total kidney volume progression over time and clearly detect kidney cysts. Since many ADPKD patients also have polycystic livers, total liver and liver cyst volume may provide additional relevant information. Total kidney, liver, and cyst volume measurements are based on kidney segmentation, which is generally performed by manual contouring, an operator-dependent and time-consuming task requiring dedicated expertise. Automatic or semi-automatic methods for kidney and cysts segmentation have been proposed in the past based on traditional approaches and artificial intelligence (AI) techniques. However, the low explainability and the need of large and curated datasets allowing to obtain accurate and generalized models have so far hampered their wide adoption in clinical research. A fully automatic segmentation method for the accurate identification of kidneys, liver, and cysts would be highly desirable. Beyond kidney, liver, and cyst volume quantification, the characterization of non-cystic renal tissue may provide additional relevant information on ADPKD pathophysiology. Few years ago, a contrastenhanced CT study revealed the presence of peritubular interstitial fibrosis in the non-cystic component of ADPKD kidneys, that was associated with renal function and its decline over time, confirmed more recently by an independent study on dynamic contrast-enhanced T1-weighted MRI. Advanced image processing techniques, such as radiomics, which aims to compute high throughput information from radiological images for the characterization of tissue spatial heterogeneity, show potential to characterise tissue microstructure. Preliminary attempts on ADPKD patients were performed on T1-weighted and T2-weighted MRI scans. Besides, radiomics could be helpful to build multiparametric stratification and prediction models including image-based features.