Pre‐operative prediction of lymph node metastasis (LNM) in patients with colorectal cancer (CRC) is challenging, yet crucial for prognosis and treatment. This study aimed to develop and validate a computed tomography‐based radiomics model to predict LNM preoperatively and to investigate its prognostic value.

A total of 587 individuals with histologically confirmed CRC from two medical centers were retrospectively analyzed. From these, 257, 109, and 221 were allocated to training, internal validation, and external validation cohorts, respectively. A total of 1781 radiomics features were obtained from portal venous‐phase computed tomography images. After feature selection, five machine learning classifiers were developed and compared. The optimal radiomics model was integrated with significant clinical predictors to develop a combined model whose performance was evaluated using receiver operating characteristics, calibration, and decision curves. The model's prognostic value was determined using Kaplan–Meier curve and Cox regression analyses.

Among the models, an extreme gradient boosting classifier demonstrated the best performance, achieving area under the receiver operating characteristic curves of 0.826, 0.807, and 0.752 in the training, internal validation, and external validation cohorts, respectively. The combined model integrating radiomics features and carcinoembryonic antigen levels showed higher predictive value, with area under the receiver operating characteristic curves of 0.842, 0.812, and 0.770 for the three cohorts. Risk stratification based on the combined model effectively identified patients with significantly different overall survival and disease‐free survival (log‐rank test, all p < 0.05). The model remained an independent predictor of both disease‐free survival (hazard ratio = 2.857, 95% confidence interval: 1.694–4.818) and overall survival (hazard ratio = 1.975, 95% confidence interval: 1.001–3.919) in multivariable Cox analysis.

Our proposed radiomics‐based model demonstrated good performance in preoperative prediction of LNM and could provide valuable prognostic information for patients with CRC.

With the advance in digital pathology and artificial intelligence (AI)-powered approaches, necrosis is proposed as a marker of poor prognosis in colorectal cancer (CRC). However, most previous studies quantified necrosis merely as a tissue type and patch-level segmentation. Thus, it was worth exploring and validating the prognostic and predictive value of necrosis proportion with a pixel-level segmentation in large multicenter cohorts.

A semantic segmentation model was trained with 12 tissue types labeled by pathologists. Segmentation was performed using the U-net model with a subsequently derived necrosis tumor ratio (NTR). We proposed the NTR score (NTR-low or NTR-high) to evaluate the prognostic and predictive value of necrosis for disease-free survival (DFS) and overall survival (OS) in the development (N = 443) and validation cohorts (N = 333) using 75% as a threshold.

The 2-category NTR was an independent prognostic factor and NTR-low was associated with significant prolonged DFS (unadjusted HR for high vs. low 1.72 [95% CI 1.19–2.49] and 1.98 [1.22–3.23] in the development and validation cohorts). Similar trends were observed for OS. The prognostic value of NTR was maintained in the multivariate analysis for both cohorts. Furthermore, a stratified analysis showed that NTR-high was a high risk with adjuvant chemotherapy for OS in stage Ⅱ CRC (p = 0.047).

AI-based pixel-level quantified NTR has a stable prognostic value in CRC associated with unfavorable survival. Additionally, adjuvant chemotherapy provided survival benefits for patients with a high NTR score in stage Ⅱ CRC.