37 - Development and Preliminary Validation of a Radiomics and Deep Learning Model for Predicting Perioperative TIPS Portal Pressure Gradient (PPG)

Purpose: Transjugular intrahepatic portosystemic shunt (TIPS) is an effective therapeutic procedure for portal hypertension, but its clinical application is limited by postoperative complications such as overt hepatic encephalopathy (OHE). The portal pressure gradient (PPG) is crucial for the efficacy and safety of TIPS. However, current PPG measurement can only be performed during invasive interventional procedures, and there is a lack of preoperative prediction methods. This study aims to develop a non-invasive tool integrating deep learning (DL) and radiomics for the PPG in the perioperative period of TIPS, including preoperative PPG (pre - PPG) and postoperative PPG (post - PPG). Meanwhile, it intends to establish a risk stratification for OHE after TIPS based on the predicted post - PPG.

Materials and Methods: A multicenter retrospective study enrolled 392 patients with portal hypertension (PH) who underwent TIPS. 306 patients were from Nanfang Hospital and randomly divided into training set and internal test set, while 86 patients were from other hospitals serving as the external test set. A semi-supervised model was used for automatic 3D dimensional segmentation of the portal vein on enhanced CT obtained before TIPS. Radiomics and deep learning features were extracted. After comparing multiple machine learning models, a MLP model was finally selected and constructed to predict the pre and post PPG of TIPS. Based on the results of predicted post-PPG , the Maximally Selected Log-rank Statistic was used to establish the p-PPG risk stratification (P-PPG-RS), and its performance was validated through K-M curves, Cox regression, and DCA. Eventually, a software was developed to facilitate its promotion and application in clinical practice.


Results:


The semi-supervised segmentation model achieved a Dice Similarity Coefficient of 0.9436, with segmentation time reduced from 20 minutes to 2 seconds.The MLP model combining radiomics and deep learning features exhibited the optimal predictive performance in both pre and post - PPG prediction. A threshold of 9.0319 in the post-PPG effectively stratified patients into high and low OHE risk groups. Integration of P-PPG-RS improved the C-index of Cox models and showed higher net benefit than Child-Pugh and MELD scores in DCA.










Conclusion: The hybrid model integrating DL and radiomics features demonstrates high accuracy in predicting perioperative TIPS PPG. Additionally, the P-PPG-RS for OHE established in this study enables reliable identification of patients at high or low risk of post-TIPS OHE.