LRFNet: A deep learning model for the assessment of liver reserve function based on Child-Pugh score and CT image

作者：Huang, Zhiwei; Zhang, Guo; Liu, Jiong; Huang, Mengping; Zhong, Lisha^*; Shu, Jian^*

来源：Computer Methods and Programs in Biomedicine, 2022, 223: 106993.

DOI：10.1016/j.cmpb.2022.106993

• Liverreservefunction
• Deeplearning
• Computedtomography
• Child-Pughscore
• Medicalimageanalysis

Background and objective: Liver reserve function should be accurately evaluated in patients with hepatic cellular cancer before surgery to evaluate the degree of liver tolerance to surgical methods. Meanwhile, liver reserve function is also an important indicator for disease analysis and prognosis of patients. Child -Pugh score is the most widely used liver reserve function evaluation and scoring system. However, this method also has many shortcomings such as poor accuracy and subjective factors. To achieve comprehen-sive evaluation of liver reserve function, we developed a deep learning model to fuse bimodal features of Child-Pugh score and computed tomography (CT) image. Methods: 1022 enhanced abdomen CT images of 121 patients with hepatocellular carcinoma and im-paired liver reserve function were retrospectively collected. Firstly, CT images were pre-processed by de -noising, data amplification and normalization. Then, new branches were added between the dense blocks of the DenseNet structure, and the center clipping operation was introduced to obtain a lightweight deep learning model liver reserve function network (LRFNet) with rich liver scale features. LRFNet extracted depth features related to liver reserve function from CT images. Finally, the extracted features are input into a deep learning classifier composed of fully connected layers to classify CT images into Child-Pugh A, B and C. Precision, Specificity, Sensitivity, and Area Under Curve are used to evaluate the performance of the model. Results: The AUC by our LRFNet model based on CT image for Child-Pugh A, B and C classification of liver reserve function was 0.834, 0.649 and 0.876, respectively, and with an average AUC of 0.774, which was better than the traditional clinical subjective Child-Pugh classification method. Conclusion: Deep learning model based on CT images can accurately classify Child-Pugh grade of liver reserve function in hepatocellular carcinoma patients, provide a comprehensive method for clinicians to assess liver reserve function before surgery.

• 单位
  西南医科大学