Deep learning or radiomics based on CT for predicting the response of gastric cancer to neoadjuvant chemotherapy: a meta-analysis and systematic review

Zhixian Bao ^1,2, Jie Du ³, Ya Zheng ^1,4

⁰Department of Gastroenterology, the First Hospital of Lanzhou University, Lanzhou, China.

Frontiers in Oncology +

|

April 11, 2024

View abstract on PubMed

Summary

Artificial intelligence (AI) models show promise in predicting neoadjuvant chemotherapy response in gastric cancer patients. While integrated models offer higher sensitivity, AI models demonstrate comparable specificity, highlighting AI

Area Of Science

Oncology
Medical Imaging
Artificial Intelligence

Background

This study evaluates the accuracy of artificial intelligence (AI) models, clinical models (CM), and integrated models (IM) for predicting neoadjuvant chemotherapy (NACT) response in gastric cancer (GC) patients.
The research aims to identify the diagnostic capabilities of AI models and compare their performance against CM and IM using head-to-head comparative studies.

Approach

A systematic literature search was conducted across PubMed, Web of Science, Cochrane Library, and Embase until September 5, 2023.
Nine studies involving 3313 patients were included for AI model analysis, with seven head-to-head studies comprising 2699 patients.
Study quality was assessed using QUADAS-2 criteria, and diagnostic accuracy was analyzed using forest plots, HSROC curves, meta-regression, and Deeks' funnel plots.

Key Points

AI models demonstrated a pooled sensitivity (SEN) of 0.75 and specificity (SPE) of 0.77 across nine studies.
In head-to-head comparisons, AI models showed pooled SEN of 0.77 and SPE of 0.79.
The integrated model (IM) exhibited higher pooled SEN (0.83) but similar SPE (0.69) compared to AI models, with no significant statistical difference noted in pairwise comparisons.
Subgroup analyses indicated that factors like cohort type, cut-off value, and gold standard choice influence model performance.

Conclusions

AI models are viable for predicting NACT response in GC patients, with CT-based deep learning models showing particular sensitivity.
While IMs offer higher sensitivity, AI models provide comparable specificity, suggesting their potential utility in clinical decision-making.
Further large-scale, rigorously designed diagnostic accuracy and head-to-head comparative studies are recommended to validate these findings.

Keywords:

artificial intelligence deep learning gastric cancer meta-analysis neoadjuvant chemotherapy radiomics