Project Details
Projekt Print View

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

Applicant Dr. Lukas Müller
Subject Area Radiology
Term from 2023 to 2024
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 518477942
 
Final Report Year 2024

Final Report Abstract

Hepatocellular carcinoma (HCC) is the most common primary liver malignancy, occurring in over 80% of cases as a result of liver cirrhosis. This creates a dualism of tumor and chronic disease that mutually influence each other. Treatment decisions are challenging, as tumor burden and liver function alone are often insufficient for prognosis. Additionally, morphological changes and portal hypertension arise, further limiting treatment options. Today non-invasively assessable surrogate parameters for estimation of the portal pressure are available. Previous work has shown that the mere presence of ascites strongly correlates with prognosis, but the relevance of ascites volume remains unclear. Manual volume quantification is too labor-intensive, but automated volumetry using deep learning techniques could efficiently capture this important prognostic factor. Therefore, this project aimed to investigate the prognostic role of quantified ascites volume in patients with HCC and develop an algorithm for automated ascites quantification. By the end of the funding period, manual segmentation of 125 CT datasets from patients with HCC undergoing TACE or immunotherapy had been completed. Two publications demonstrated that ascites is a highly relevant prognostic factor for patients with HCC, with the manual segmentation results indicating that ascites volume plays a significant role. The automated quantification algorithm developed during the project achieved a DICE score of 0.84 in the training dataset (n=101 patients with HCC and TACE therapy) and 0.74 in the validation set (n=24 patients with HCC and immunotherapy), which is currently considered clinically insufficient by the applicants. Therefore, further improvement in training and validation is planned after the funding period, with the goal of integrating the algorithm into broader predictive models. In summary, our results show that quantifying ascites provides important prognostic information and could play a role in non-invasive assessment of portal hypertension in the future. Due to its complexity, manual quantification is not feasible in clinical practice, making an automated solution necessary. The developed algorithm currently shows only moderate accuracy, and further improvements are planned to enable clinical use.

Publications

 
 

Additional Information

Textvergrößerung und Kontrastanpassung