990P - Prognosis prediction for unresectable hepatocellular carcinoma undergoing hepatic arterial infusion chemotherapy using convolutional neural network model

Background

Hepatic arterial infusion chemotherapy (HAIC) is known to be more effective than conventional systemic chemotherapy, showing great potential in treating unresectable hepatocellular carcinoma (HCC) patients. However, there is still unclear which group can benefit more from HAIC.

Methods

191 patients with unresectable HCC undergoing HAIC from Zhongshan Hospital between May 2019 and March 2022 were retrospectively recruited. Radiomics scores were calculated based on enhanced-T1-weighted, enhanced-T2-weighted, arterial phase and delayed phase images. Clinical factors related to OS and PFS were identified by Cox regression analysis. Three different CNNs of AlexNet, ResNet, and Inception architectures were constructed on 70% of the data set and tested on the remaining 30%. Radiomics scores and clinical factors were reflected to a model eventually. Mean squared error (MSE) and time-dependent receiver operating characteristic curve were calculated for the models.

Results

The OS model included radiomics score with No. of HAIC cycles, tumor thrombus, PIVKA-II, neutrophil-lymphocyte ratio (NLR), aspartate aminotransferase (AST), gamma-glutamyltranspeptidase (γ-GT) and C-reactive protein. And the PFS model included radiomics score with No. of HAIC cycles, tumor thrombus, NLR and γ-GT. The AlexNet-OS model, ResNet-OS model, and Inception-OS model achieved the best MSE of 1.0068, 0.9023 and 0.8506, respectively. The AlexNet-PFS model, ResNet- PFS model, and Inception- PFS model achieved the best MSE of 0.6658, 0.6819 and 1.1012, respectively.

Conclusions

The present models which integrated radiomics information and clinical factors helped predict OS and PFS of unresectable HCC patients undergoing HAIC treatment.

Clinical trial identification

Editorial acknowledgement

We are very grateful to Mrs. Hailin Mi for helping us to construct the deep learning model. Hailin Mi kindly provided statistical advice for this manuscript. (Hailin Mi; Department of Computer Science and Technology, Harbin Engineering University, China)

Legal entity responsible for the study

The authors.

Funding

The National Natural Science Foundation of China (No. 81972889, Dr Yin).

Disclosure

All authors have declared no conflicts of interest.