A Multi-Organ Fusion and LightGBM Based Radiomics Algorithm for High-Risk Esophageal Varices Prediction in Cirrhotic Patients

Abstract

Esophageal varices (EV) is the most common complication of portal hypertension in cirrhosis patients. Radiomics has been progressing remarkably for quantifying the state of diseases. However, there are few studies on EV severity prediction by applying radiomics and machine learning. Besides, most of the existing methods apply only single organ for radiomics feature extraction. In this study, we propose a radiomics algorithm based on light gradient boosting machine (LightGBM) to identify high-risk and lowrisk EV patients by fusing the radiomics features of liver, spleen and esophagus from the CT images. This approach involves 188 patients, including 151 cirrhotic patients (84 patients with severe EV and 67 patients with mild or no EV) registered in Qilu Hospital of Shandong University and 37 cirrhotic patients (20 patients with severe EV and 17 patients with mild or no EV) retrospectively registered in Jinan Central Hospital from January 2018 to August 2020. Specifically, the radiomics features of liver, spleen and esophagus are extracted after manual segmentation. Then the features of the three organs are fused by linear combination where the weights are estimated by the feature distribution. Finally, classification models are established by cross-combination of multiple feature selection methods (e.g., least absolute shrinkage and selection operator, Boruta, eXtreme gradient boosting (XGBoost) and LightGBM) with multiple classifiers (e.g., support vector machine, random forests, XGBoost and LightGBM) to discriminate the high-risk EV patients from the low-risk EV patients at the individual level. In addition, we propose a classifier ensemble strategy by combining the prediction probability of each organ for final classification. Experimental results demonstrate that the feature of esophageal makes a greater contribution on the diagnosis of EV compared with that of spleen and liver. The proposed multi-organ fusion and LightGBM based radiomics framework has better classification performance compared with the state-of-the-art radiomics approaches.INDEX TERMS Radiomics, esophageal varices, multi-organ feature fusion, classifier ensemble, LightGBM.

Study on the Drivers of Inclusive Green Growth in China Based on the Digital Economy Represented by the Internet of Things (IoT)

“…In feature selection, we only need to find the optimal segmentation point according to the discrete value of the histogram. The leaf-wise strategy with depth limitation saves a lot of time and space [ 61 ].…”

Evaluating the Performance of Inclusive Growth Based on the BP Neural Network and Machine Learning Approach

“…In feature selection, we only need to traverse to find the optimal segmentation point according to the discrete value of the histogram. In addition, the use of leaf wire strategy with depth limit saves a lot of time and space consumption [ 49 , 50 ]. LightGBM incorporates many T regress tress ∑ t =1 T g t ( X ) to estimate the final model which may be expressed as follows: where the regress trees may be expressed as W p ( x ) , and p ∈ {1,2,…, N }.…”

Applying an Improved Stacking Ensemble Model to Predict the Mortality of ICU Patients with Heart Failure

“…This process greatly accelerates the prediction speed and reduces storage space required without degrading the prediction accuracy [ 58 , 59 ]. LGBM is a gradient boosting decision tree learning algorithm that has been widely used for feature selection, classification and regression [ 60 ].…”