Combining radiomics and deep learning to predict liver metastasis of gastric cancer on CT image

Abstract

Objective: Our study aimed to explore the potential of deep learning (DL) radiomics features from CT images of primary gastric cancer (GC) in predicting gastric cancer liver metastasis (GCLM) by establishing and verifying a prediction model based on clinical factors, classical radiomics and DL features.

Methods: We retrospectively analyzed 1001 pathologically confirmed GC patients from June 2014 to May 2024, divided into non-LM (n=689) and LM groups (n=312). CT-based classic radiomics and DL features were extracted and screened to construct a DL-radiomics score. This score, along with statistically significant clinical factors, was used to build a fused model which visualized as a nomogram. The model's predictive performance, calibration, and clinical utility were assessed and compared against a clinical model. Additionally, the DL-radiomics score's role in distinguishing between synchronous and metachronous GCLM was evaluated.

Results: The fused model showed good predictive performance [AUC: 0.796 (95% CI: 0.766-0.826) in training cohort and 0.787 (95% CI: 0.741-0.834) in test cohort], outperforming the clinical model, radiomics score and DL score (P<0.05). In addition, the decision curve confirmed that the model provided the largest clinical net benefit compared with all other models in the relevant threshold. DL-radiomics score showed moderate predictive performance in distinguishing between synchronous GCLM and metachronous GCLM, with an AUC of 0.665 (95% CI, 0.613-0.718).

Conclusion: The CT-based fused model has demonstrated significant value in predicting the occurrence of GCLM, and can provide a reference for the personalized follow-up and treatment of patients.

Keywords: computed tomography; deep learning; gastric cancer; liver metastasis; radiomics nomogram.

Figure 1

Flowchart of patient selection process. GC, gastric cancer; LM, liver metastases; GCLM, gastric cancer liver metastases.

Figure 2

Overview of the study design. (A) Collection of 5mm venous phase CT Images; (B) Extraction of DL features and classical radiomics features; (C) Feature selection and model construction; (D) Model visualization and evaluation. DL, deep learning.

Figure 3

Comparison of different models. ROC curves of different models to predict the occurrence of GCLM, in training cohort (A) and test cohort (C); the heat map shows that the DeLong test compares the statistical results of the AUC values of different models, in training cohort (B) and test cohort (D). DL, deep learning; ROC, receiver operator characteristic; GCLM, gastric cancer liver metastases.

Figure 4

The violin plots showing distribution of different radiomics scores between Non-LM group and LM group in training cohort. (A) Radiomics score; (B) DL score; (C) DL-radiomics score. LM, liver metastases; DL: deep learning; LM: liver metastases.

Figure 5

Fused nomogram with the DL-radiomics score and clinical factors (tumor thickness, CEA and CA199). DL, deep learning; CEA, carcinoembryonic antigen; CA199, Carbohydrate antigen199.

Figure 6

Decision curves analysis for different models. DL, deep learning.

Figure 7

Calibration curve of Fused model to predict the of GCLM occurrence. GCLM, gastric cancer liver metastases.

Figure 8

The violin plot illustrating the distribution of DL-Radiomics score for both synchronous GCLM and metachronous GCLM. DL, deep learning; GCLM, gastric cancer liver metastases.