Patient characteristics

The characteristics of the 629 enrolled patients are shown in Table 1. The median follow-up time for distant metastasis-free survival (DMFS) were 49.6 months in the primary cohort (interquartile range [IQR]: 47.3–52.6 months), 52.6 months in validation cohort 1 (IQR: 50.3–58.9 months), 43.1 months in validation cohort 2 (IQR: 42.3–45.7 months), and 46.3 months in validation cohort 3 (IQR: 45.2–48.2 months).

Satisfactory inter- and intra-observer reproducibility were observed for the tumor masking and radiomic feature extraction (ICC > 0.6)²² when we compared results for five radiologists and results from the same radiologist at baseline and at least 1 month later.

Radiomic signature construction and validation

The coarse-to-fine feature selection strategy identified four relevant features (Supplementary Table 1). The selected features were incorporated into a least absolute shrinkage and selection operator (LASSO)-Cox regression model to define the radiomic signature. For each of the primary cohort and the three validation cohorts, patients were classified into high- and low-radiomic signature groups for further analyses based on the median radiomic signature value of the primary cohort. The Kaplan–Meier survival curves confirmed a significant difference in DMFS between the high- and low-radiomic signature groups (p < 0.001), with relatively high hazard ratios (HRs, >3.9) in all four cohorts (Fig. 1a–d, upper). The areas under the curve (AUCs) at different follow-up times (1, 2, and 3 years) also confirmed that the radiomic signature had good prognostic accuracy in the primary and validation cohorts (Fig. 1a–d, lower). Subgroup analyses further confirmed that the radiomic signature could predict prognosis according to clinical stage (Fig. 2) as well as in the pT and pN subgroups from each cohort (Supplementary Figs. 1 and 2). These results confirmed the high prognostic accuracy of the radiomic signature.

Incremental value of the radiomic signature

Multivariate Cox analysis revealed that DM was independently predicted by the radiomic signature, surgery location, and pN stage. Therefore, a radiomic nomogram (Fig. 3a) and clinical models (Supplementary Fig. 3) were constructed using the primary cohort. The calibration curves for the radiomic nomogram at 1 year, 2 years, and 3 years showed good agreement between the estimations and the clinical outcomes in the primary and validation cohorts. The C-index values for the different models, namely radiomic signature, radiomic nomogram, clinical nomogram, and VN, are listed in Table 2. Relative to the clinical nomogram and the VN, the radiomic signature provided better performance in the primary cohort (C-index: 0.847, 95% confidence interval [CI]: 0.803–0.891) and the validation cohorts (validation cohort 1: C-index: 0.809, 95% CI: 0.718–0.901; validation cohort 2: C-index: 0.848, 95% CI: 0.761–0.934; validation cohort 3: C-index: 0.803, 95% CI: 0.705–0.901) (Table 2). Furthermore, the radiomic nomogram based on the radiomic signature and clinicopathologic factors (Supplementary Table 2) also achieved better performance and significantly improved the classification accuracy for DMFS outcomes, based on the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) values (Supplementary Fig. 4).

The decision curve analysis revealed that the radiomic nomogram had relatively good clinical performance, with advantages across almost the entire range of reasonable threshold probabilities in the primary and validation cohorts.

These results suggested that the radiomic signature provided additional value for personalized DM prediction.

Risk stratification using the radiomic signature

In order to detect patients that can benefit from adjuvant chemotherapy, interaction tests among radiomic signature, pathological stage, and adjuvant chemotherapy efficacy were performed (Table 3).

The interaction test for radiomic signature and adjuvant chemotherapy efficacy revealed that the adjuvant chemotherapy benefit was worse among patients with a high-radiomic signature (HR: 1.706, 95% CI: 1.131–2.572, p < 0.05; p < 0.001 for interaction), relative to among patients with a low-radiomic signature. The corresponding Kaplan–Meier DMFS curves are shown for the high- and low-radiomic signature groups in Fig. 4a. Adjuvant chemotherapy was significantly associated with decreased DMFS in the high-radiomic signature group (p = 0.01), did not have a significant association in the low-radiomic signature group, and had only a marginally significant association among all patients (p = 0.087) (Fig. 4a). These results suggest that LARC patients with a high-radiomic signature may experience even worse outcomes after receiving adjuvant chemotherapy.

The interaction tests for radiomic signature and pathological stage revealed that both pT stage and pN stage were associated with DMFS among all patients (Fig. 4b, c). Advanced stages suggested high risk of DM, meaning that patients with higher stage usually showed decreased DMFS. Specifically, pT stage was significantly associated with DMFS in the high-radiomic signature group (HR: 1.620, 95% CI: 1.140–2.303, p < 0.05) not in the low-radiomic signature group, and pN stage was significantly associated with DMFS in both the high- and low-radiomic signature groups (HR: 1.904, 95% CI: 1.526–2.375, p < 0.05 in high-radiomic signature group, and HR: 2.108, 95% CI: 1.141–3.895, p < 0.05 in low-radiomic signature group).

The interaction tests for pathological stage and adjuvant chemotherapy efficacy in the high- and low-radiomic signature groups were also performed. The results for pT stage subgroup analysis indicated that, in the high-radiomic signature group, pT1–2 patients did not benefit from the adjuvant chemotherapy (HR: 11.661, 95% CI: 1.531–88.825, p = 0.003; p < 0.001 for interaction), while no significant interactions were observed in the low-radiomic signature group (Fig. 5). The results for pN stage subgroup analysis indicated that, pN0 patients with high-radiomic signature and adjuvant chemotherapy, had even worse survival than those with high-radiomic signature but without adjuvant chemotherapy (HR: 2.666, 95% CI: 1.269–5.601, p = 0.007; p < 0.001 for interaction), while in the low-radiomic signature group, only pN2 patients had survival benefit from the adjuvant chemotherapy (HR: 0.177, 95% CI: 0.029–1.064, p = 0.033; p < 0.001 for interaction) (Fig. 6).

These results of interaction tests suggest that not all LARC patients will benefit from adjuvant chemotherapy, and the treatment strategy should be carefully selected based on the pathological stage and radiomic signature as well.