Deep learning model based on MRI-derived microvascular network simulation parameters for noninvasive assessment of lymphovascular invasion in rectal cancer patients.

Background: Accurate preoperative assessment of lymphovascular invasion (LVI) in patients with rectal cancer (RC) is important for guiding postoperative management. This study aimed to develop and validate a deep learning model based on magnetic resonance imaging (MRI)-derived microvascular network simulation parameters for the preoperative assessment of LVI in RC patients. Methods: A total of 453 patients with pathologically confirmed rectal adenocarcinoma from two medical centers were retrospectively enrolled. All patients underwent multi-b-value diffusion-weighted imaging (DWI) before surgery. First, a steady-state Navier-Stokes hemodynamic model of the tumor microvascular network was constructed based on the multi-b-value DWI images. Subsequently, voxel-wise least squares fitting was performed to match the DWI signals with the dictionary, enabling the inversion and generation of spatial parametric maps for mean flow velocity (V-m), velocity standard deviation (V-s), and angiogenic branching index (ANB). These parametric maps were then input into a Vision Transformer (ViT) network to extract deep features from each modality. A cross-attention fusion module was designed to capture spatial interactions among the parametric maps and construct a multiparametric fusion model. The model's performance was comprehensively evaluated using the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Results: The multiparametric fusion model achieved favorable performance, with AUCs of 0.901 [95% confidence interval (CI): 0.808-0.993] and 0.863 (95% CI: 0.800-0.926) in the internal and external validation cohorts, respectively. DCA demonstrated that within the threshold range of 0.2-0.8, the fusion model provided substantially greater clinical net benefit than the individual parameter models. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization revealed that the model's attention was primarily focused on the invasive front of the tumor and regions with high peritumoral vascular density, providing supportive visual evidence and suggesting potential biological relevance. Conclusions: The deep learning model based on MRI-simulated microvascular network parameters provides a promising and noninvasive approach for the preoperative assessment of LVI status in RC patients. The model demonstrated encouraging performance in both internal and external validation cohorts. However, further prospective and multicenter validation is required before clinical application.