In the present study, we proposed and validated an effective DL signature based on pretherapy CT images for TRG discrimination in LAGC patients treated with NCT. Additionally, the added values to identified clinicopathological characteristics for TRG and OS predictions by MLR and MCR methods were verified.
NCT is an important treatment for LAGC, but different patients have different responses to NCT. At present, there is no reliable and effective method for predicting the efficacy of NCT for LAGC, which leads to the failure of NCT in some patients with LAGC, and some patients even miss the chance of radical surgery due to disease progression during chemotherapy. Therefore, the development of a model to accurately predict the efficacy of NCT before treatment is of great significance for the precise treatment of LAGC patients.
In recent years, increasing attention has been given to the study of radiomics because radiomics can extract and analyze a large number of advanced quantitative imaging features that could reflect the heterogeneity of the tumor. Radiomic features have clinical value in the early prediction and identification of patients who are sensitive to NCT. An earlier study found that the radiomics features screened by CT imaging before treatment were important markers of response to NCT in LAGC[27]. Peng et al.[28] performed radiomic feature extraction on the portal vein CT images of 106 GC patients before NCT and established an efficacy prediction model of NCT using a random forest algorithm, which showed good predictive performance in the validation cohort, with an AUC of 0.82. Zhou et al.[10] extracted radiomic features from the CT images of 323 GC patients and observed that the radiomics signature had good discrimination performance for predicting NCT response in the external cohort (AUC, 0.679; 95% CI, 0.554–0.803). In addition, a radiomic model for predicting the efficacy of NCT in GC was constructed using a Bayesian classifier, support vector machine, random forest and other algorithms, and good discrimination performance was observed in both the internal validation cohort (AUC, 0.784; 95% CI, 0.659–0.908) and external validation cohort (AUC, 0.803; 95% CI, 0.717–0.888)[29]. However, the clinical relevance of these findings is limited because of the relatively small sample sizes of the studies and lack of validation in multicenter cohorts. In the latest, the process of DL radiomic feature extraction was performed in a larger population (719 patients) for predicting the efficacy of NCT in GC, and higher AUCs of 0.804–0.829 were observed in the internal validation and external validation cohorts, whereas it lacked the end-to-end architecture of TRG prediction. Therefore, we proposed an end-to-end DL model to extract richer information in larger and wider datasets.
The size of the GR group is usually several times larger than that of the PR group. Achieving a balanced imaging dataset is either resource-consuming or lacks algorithm complexity, which causes unstable results, especially when a large number of images need to be generated. Accordingly, we used a state-of-the-art oversampling algorithm, DeepSMOTE, to enrich the information in the GR group in an attempt to generate more GR images to improve the discriminative performance of the model. With the visualized output map of the tumor area, revealing the imaging characteristics extracted by the DL signature connected to the intratumor heterogeneity may contribute to TRG prediction in GC. From the heatmaps, we found that most of the important features for making decision could be caught by the reconstructed tumors as well. Moreover, we explored the effectiveness of the DL signature in predicting OS. Our previous study found that the TRG score was related to LAGC patient prognosis after D2 gastrectomy[3]. In this study, we found that the DL signature was an independent risk factor for survival in LAGC patients treated with NCT. Patients with higher DL scores were indicated to have better OS. More specifically, patients with GR after NCT could benefit greatly in terms of survival. We also found that low differentiation, Borrmann type IV, high level of CEA before NCT and cN stage were independent risk factors for survival in LAGC patients, which is similar to the findings of many other previous studies. Therefore, the proposed nomogram may provide a feasible way to guide treatment plans and implement personalized treatment for LAGC patients treated with NCT.
However, there are some limitations in this study. First, as a retrospective and multicenter study, there was possible selection bias and inherent bias. For example, patients from different levels of hospitals using different CT devices may cause bias. Therefore, it is necessary to design prospective studies to validate the generalizability and clinical applicability of our models. Second, although we visualized the intratumor characteristics extracted by the DL signature, its clear biological significance is still unknown and needs to be fully elucidated. Further exploration of the relationship between radiographic features and the tumor microenvironment may provide more micro information and biological significance of DL. Third, the extracted imaging features are closely dependent on the ROIs; however, the precise manually delineated tumor margins require professional expertise, which is highly influenced by subjective experience. Therefore, an automated tumor segmentation mechanism for CT images in GC needs to be further developed for more precise TRG prediction. Fourth, validation of the clinical reliability of the images generated by DeepSMOTE in the GR group should be explored in future studies since the algorithm has only been widely assessed on natural images.