AI-based ascites detection and quantification in patients with hepatocellular carcinoma - development, translation and clinical relevance

Hepatocellular carcinoma (HCC) is the most common primary liver malignancy. More than 80% of patients develop HCC as a consequence of liver cirrhosis. Thus, there is often a dualism of tumor manifestation and chronic disease, which mutually potentiate and influence each other's course. In addition, treatment decisions are complicated because tumor burden alone is only partially sufficient to determine the prognosis of patients. However, even when including laboratory parameters that correlate with the remaining liver synthesis capacity, therapy decisions remain difficult. This is due to the fact that, in addition to a reduction in liver synthesis capacity, morphological changes in the liver and portal venous inflow occur in the context of chronic inflammation. In the majority of patients, these lead to portal venous hypertension, which in turn is associated with a restriction of the treatment ability of the patients and thus significantly influences their prognosis. The gold standard for the determination of portal venous pressure is a direct transjugular measurement, which is not a standard in the initial diagnostic pathway in patients with HCC due to its invasiveness and associated risks. Several surrogate parameters for noninvasive determination of portal venous pressure from image data have been proposed. In a preliminary work, we could show that the simple presence of ascites has a high correlation with the prognosis of the patients when determined qualitatively in binary manner. However, there is no evidence on how relevant the quantified ascites volume is for the prognosis of the patients, although significant inter-individual differences are present and volumes range from a few milliliters to several liters. The main problem is the laborious manual ascites quantification from cross-sectional image data, which cannot be reasonably integrated into the daily radiological routine. However, automated volumetry using deep learning techniques could perform this task without any loss of time. This would automatically provide an additional, highly relevant prognostic factor in each radiology report. The aim of this project is to investigate the role of quantified ascites volume in patients with HCC. In a first step, an algorithm for fully automated ascites determination from routine data sets will be trained and validated. In a second step, the ascites volume will be correlated with prognostic parameters and the risk assessment of hepatic decompensation under therapy. Finally, a model for improved prognosis determination in patients with HCC will be developed using this information in combination with previously known risk factors.