Advanced gastric cancer: CT radiomics prediction and early detection of downstaging with neoadjuvant chemotherapy

Objectives To develop and evaluate machine learning models using baseline and restaging computed tomography (CT) for predicting and early detecting pathological downstaging (pDS) with neoadjuvant chemotherapy in advanced gastric cancer (AGC). Methods We collected 292 AGC patients who received neoadjuvant chemotherapy. They were classified into (a) primary cohort (206 patients with 3-4 cycles chemotherapy) for model development and internal validation, (b) testing cohort I (46 patients with 3-4 cycles chemotherapy) for evaluating models' predictive ability before and after the complete course, and (c) testing cohort II (n = 40) for model evaluation on its performance at early treatment. We extracted 1,231 radiomics features from venous phase CT at baseline and restaging. We selected radiomics models based on 28 cross-combination models and measured the areas under the curve (AUC). Our prediction radiomics (PR) model is designed to predict pDS outcomes using baseline CT. Detection radiomics (DR) model is applied to restaging CT for early pDS detection. Results PR model achieved promising outcomes in two testing cohorts (AUC 0.750, p = .009 and AUC 0.889, p = .000). DR model also showed a good predictive ability (AUC 0.922, p = .000 and AUC 0.850, p = .000), outperforming the commonly used RECIST method (NRI 39.5% and NRI 35.4%). Furthermore, the improved DR model with averaging outcome scores of PR and DR models showed boosted results in two testing cohorts (AUC 0.961, p = .000 and AUC 0.921, p = .000). Conclusions CT-based radiomics models perform well on prediction and early detection tasks of pDS and can potentially assist surgical decision-making in AGC patients.

Automated summary

The study developed and evaluated machine learning models using baseline and restaging CT for predicting and early detecting pathological downstaging in AGC patients receiving neoadjuvant chemotherapy. The radiomics models performed well on prediction and early detection tasks, potentially assisting surgical decision-making for AGC patients.