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

Introduction

Rectal cancer (RC) accounts for approximately 30–40% of all colorectal cancer cases, with nearly 20% of patients presenting with locally advanced disease at initial diagnosis (1). Lymphovascular invasion (LVI) defined as the penetration of tumor cells through the intestinal wall into venous or lymphatic vessels, is a critical pathological feature in the progression of RC (2). Both clinical and pathological studies (3,4) have demonstrated that positive LVI status is significantly associated with increased postoperative recurrence rates and a higher risk of distant metastasis, serving as an independent adverse prognostic factor for long-term outcomes. The presence of LVI generally indicates that the tumor possesses the potential for hematogenous dissemination and is closely linked to tumor angiogenesis, abnormal microvascular proliferation, and remodeling of the local immune microenvironment. In addition, LVI and vascular dissemination are closely associated with nodal pathways and local lymphatic drainage patterns in RC, reflecting the anatomical and biological complexity of mesorectal lymphatic distribution (5). As an important marker of tumor aggressiveness, LVI status contributes to tumor-node-metastasis (TNM) staging and may provide additional information to assist clinical decision-making, including surgical planning and treatment strategy selection. However, treatment planning in RC is multifactorial and should not rely on a single parameter alone (6,7).

Recent advances in radiomics and deep learning have provided novel strategies for the preoperative non-invasive assessment of LVI in RC (8,9). As the cornerstone imaging modality for preoperative assessment of RC, magnetic resonance imaging (MRI), particularly T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) has been widely utilized for texture, morphological, and deep feature extraction and modeling. Although these approaches have improved diagnostic efficiency to some extent, they remain largely confined to signal-level image features. Because conventional image-based models fail to effectively characterize microvascular architecture and hemodynamic alterations within the tumor region, they lack direct biological relevance to LVI, a pathological process that strongly depends on tumor microenvironmental remodeling (10,11). To address this limitation, the present study proposes a novel preoperative assessment framework that integrates MRI-based microvascular simulation parameters with deep learning techniques. By constructing a steady-state hemodynamic-diffusion coupled model, three spatial parametric maps—mean flow velocity (V-m), velocity standard deviation (V-s), and angiogenic branching index (ANB)—were derived from multi-b-value DWI images to quantitatively depict tumor blood flow distribution and microvascular structural characteristics. This approach enables the establishment of a biologically grounded, structurally informative, and interpretable deep learning model for the non-invasive assessment of LVI in RC.

This study aimed to develop and validate a deep learning model based on MRI-derived microvascular simulation parameters for the preoperative assessment of LVI in patients with RC. By simulating physiological mechanisms, enriching structural information, and enhancing model interpretability, this work aims to extend conventional image analysis approaches and to provide more precise imaging-based support for preoperative risk stratification and personalized therapeutic decision-making. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0226/rc).

Methods

Patients

This study was a multicenter retrospective study that included a total of 453 patients with RC from two medical centers between December 2021 and January 2025. Among them, 300 patients from Center 1 (The Affiliated Huai’an Hospital of Xuzhou Medical University) were randomly divided into a training cohort (n=210) and an internal validation cohort (n=90) at a ratio of 7:3, while 153 patients from Center 2 (No. 904 Hospital of the Joint Logistics Support Force of the PLA) were included as the external validation cohort (the patient enrollment flowchart is shown in Figure 1). The inclusion criteria were as follows: (I) pathologically confirmed rectal adenocarcinoma; (II) preoperative MRI examination performed within one week before surgery with complete imaging data; and (III) complete clinical and pathological information available. The exclusion criteria were as follows: (I) received any form of preoperative treatment; (II) had other malignant tumors or metastases; and (III) poor image quality or incomplete clinical information. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of No. 904 Hospital of the Joint Logistics Support Force of the PLA (No. 20260116) and the Ethics Committee of The Affiliated Huai’an Hospital of Xuzhou Medical University (No. HEYLL20230053). The requirement for informed consent was waived because of the retrospective study design and the use of anonymized imaging and clinical data.

Figure 1 Flow chart of the patients’ recruitment pathway. Center 1, The Affiliated Huai’an Hospital of Xuzhou Medical University; Center 2, No. 904 Hospital of the Joint Logistics Support Force of the PLA. MRI, magnetic resonance imaging.

Pathological assessment

In this study, the diagnosis of LVI was based on postoperative histopathological examination. All tissue specimens were fixed, dehydrated, and embedded, followed by routine hematoxylin-eosin staining. The slides were independently reviewed by two pathologists, each with more than 5 years of experience in oncologic pathology, who were blinded to all patient information. The evaluation focused on determining whether tumor cells penetrated the basement membrane and invaded venous or lymphatic lumina, while also assessing the continuity and integrity of the vascular wall structure. To further verify suspected areas of vascular invasion, immunohistochemical staining for CD31 and CD34 was performed simultaneously to label vascular endothelial cells, thereby improving the visualization and identification of vascular structures (12). Each pathologist independently completed the diagnostic interpretation according to both cellular morphological characteristics and staining localization. In cases of disagreement, the slides were re-evaluated by a senior pathologist, whose judgment was considered final. All assessment procedures were conducted under double-blind conditions to ensure diagnostic objectivity and consistency.

MRI tumor delineation and parametric map generation

All patients underwent preoperative MRI examinations that included multi-b-value DWI sequences with b=50, 800, and 1,500 s/mm². Prior to post-processing, all DWI images were spatially registered, denoised, and intensity-normalized to ensure consistency across different b-values (detailed acquisition parameters for each scanner are provided in Table S1). The region of interest (ROI) was manually delineated by two radiologists experienced in pelvic MRI interpretation, using ITK-SNAP software, with reference to both T2WI and b =1,500 DWI images. The ROI encompassed the entire solid portion of the tumor while excluding regions of apparent necrosis or hemorrhage. The delineation was performed under double-blind conditions, and any discrepancies were resolved by consensus after review by a third senior radiologist. The inter-observer agreement for ROI delineation was excellent (Dice similarity coefficient >0.80).

Based on previously reported biophysical modeling principles of DWI signal formation (13,14), simulation-derived parametric maps were generated in this study to characterize the physiological features of the tumor microvasculature. Specifically, a previously established framework incorporating Navier-Stokes-based microvascular flow modeling and Bloch-Torrey-based diffusion-perfusion signal simulation was used to process the multi-b-value DWI data within the tumor ROI. Rather than reconstructing the hemodynamic model de novo in the current study, we implemented the published parameter-estimation pipeline to obtain voxel-wise microvascular parameter maps from the acquired DWI data. According to the established framework, the simulated signal dictionary reflects variations in microvascular flow velocity, vascular density, directionality, and branching-related characteristics under predefined model settings. The acquired DWI signals were then voxel-wise matched to the simulated dictionary using least-squares fitting to estimate the most probable microvascular functional parameters for each voxel. Finally, three microvascular network simulation parametric maps were obtained, including the V-m-map, the V-s-map, and the ANB-map. Detailed implementation of the simulation framework, including signal modeling and dictionary-based parameter estimation, as well as code availability (see Appendix 1).

Deep feature extraction and model construction

In this study, a Vision Transformer (ViT) architecture was adopted as the backbone network for feature extraction and was pretrained on the ImageNet dataset to achieve better initialization performance. The ViT network employs a self-attention mechanism to capture spatial structural information within images and has particular advantages in modeling long-range dependencies, which facilitates the preservation of the morphological and distributional characteristics of tumor regions (15). Based on the three-dimensional (3D) tumor segmentation mask, a 2.5D input image was constructed by selecting the largest tumor cross-section along with its adjacent upper and lower slices from each image modality, forming a three-channel ROI patch as the model input. The model was trained using stochastic gradient descent (SGD) optimizer with an initial learning rate of 0.001. A mini-batch size of 32 was used, and the model was trained for up to 3,000 epochs. The network was initialized with pretrained weights (ImageNet) to improve convergence. Input images were normalized using the ImageNet normalization scheme. To address class imbalance, a batch-level balancing strategy was applied during training to ensure a more balanced sampling of different classes. The training process was implemented using a standardized deep learning pipeline, and detailed implementation is provided in the publicly available code repository (URL: https://github.com/misshan2000/Deep_Learning_Transformer_Fusion). The preprocessed images were then fed into the ViT model, where spatial structural features were progressively extracted through hierarchical global self-attention layers, ultimately yielding deep feature representations that effectively characterize tumor morphology and blood flow distribution. These features were subsequently reduced to 128 dimensions using principal component analysis to minimize dimensional redundancy and reduce the risk of overfitting.

In addition, a multimodal feature fusion module based on the Transformer architecture was designed. Specifically, after spatial features were extracted from the three microvascular simulation parametric maps and the apparent diffusion coefficient map through separate ViT branches, the four sets of high-dimensional features were fed into a shared Transformer encoder for interactive learning. The fusion module consisted of three stacked Transformer encoder layers, each comprising a multi-head self-attention mechanism, a feed-forward network, and layer normalization. This structure enables the model to capture long-range dependencies and cross-feature interactions. Through a cross-channel attention mechanism, the model dynamically perceives the interrelationships among different imaging modalities, thereby enhancing its ability to represent potential patterns associated with vascular invasion. The fused feature vectors were subsequently mapped to a binary classification space through a fully connected layer, and a Sigmoid activation function was applied to output the final probability, indicating the patient’s LVI status (16) (Figure 2).

Figure 2 Workflow of model development and outcome evaluation. ADC, apparent diffusion coefficient; ANB, angiogenic branching index; DWI, diffusion-weighted imaging; MLP, multilayer perceptron; V-m, mean flow velocity; V-s, velocity standard deviation.

Statistical analysis

Statistical analyses were performed using R software (version 4.1.3; www.R-project.org) and SPSS software (version 25.0; IBM, Armonk, NY, USA). Continuous variables with a normal distribution were expressed as mean ± standard deviation, and differences between groups were evaluated using the independent-samples t-test. Continuous variables with a non-normal distribution were presented as median and interquartile range, and group comparisons were conducted using the Mann-Whitney U test. Categorical variables were summarized as counts and percentages, and between-group differences were analyzed using the chi-square test or Fisher’s exact test, as appropriate. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, and the area under the receiver operating characteristic curve. Differences in model performance were statistically compared using the DeLong test. Model calibration and goodness-of-fit were assessed using the Hosmer-Lemeshow test and calibration curves.

Results

Clinical data

Based on the inclusion and exclusion criteria, a total of 453 patients from two medical centers were retrospectively enrolled in this study between December 2021 and January 2025. Among all enrolled patients, 144 (31.8%) had LVI (LVI-positive) and 309 (68.2%) had no LVI (LVI-negative). The clinical baseline characteristics of the patients are summarized in Table 1.

Deep learning model based on ViT architecture

Based on the ViT architecture, single-parameter deep learning models were constructed separately. The area under the curve (AUC) values of the V-m-map model in the internal and external validation cohorts were 0.860 [95% confidence interval (CI): 0.754–0.965] and 0.836 (95% CI: 0.754–0.921), respectively. The V-s-map model achieved AUC values of 0.819 (95% CI: 0.703–0.935) and 0.808 (95% CI: 0.727–0.888) in the internal and external validation cohorts, respectively. The apparent diffusion coefficient (ADC)-map model yielded AUCs of 0.809 (95% CI: 0.699–0.919) and 0.793 (95% CI: 0.714–0.873) for the internal and external cohorts, respectively. The ANB-map model achieved AUCs of 0.789 (95% CI: 0.665–0.912) and 0.775 (95% CI: 0.694–0.857) in the internal and external validation cohorts, respectively. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization showed that the model’s attention was mainly focused on the invasive front of the tumor and peritumoral regions. These findings provide supportive visual evidence and suggest potential biological relevance (see Figure 3).

Figure 3 Grad-CAM heatmaps, highlighting key areas and original lesion regions predicted by the model. ADC, apparent diffusion coefficient; ANB, angiogenic branching index; Grad-CAM, Gradient-weighted Class Activation Mapping; V-m, mean flow velocity; V-s, velocity standard deviation.

Establishment and evaluation of the multi-sequence fusion model

Based on the Transformer-based feature fusion module, a multiparametric fusion model was further constructed in this study. The AUC values of the model in the internal and external testing cohorts were 0.901 (95% CI: 0.808–0.993) and 0.863 (95% CI: 0.800–0.926), respectively (Figure 4A,4B). The corresponding sensitivities were 0.931 and 0.925, and the F1-scores were 0.783 and 0.643, respectively. Compared with the single-parameter models, the fusion model showed improved performance in terms of accuracy, sensitivity, specificity, and overall assessment efficacy (Table 2). However, according to the DeLong test, the differences between the fusion model and the V-m and V-s models were not statistically significant in the internal and external validation cohorts. The Hosmer–Lemeshow test was used to assess model calibration, and the calibration curves showed good agreement between the predicted and observed outcomes, indicating satisfactory goodness of fit (P>0.05, Figure 4C,4D). The decision curve analysis (DCA) demonstrated that, across most reasonable threshold ranges, the multiparametric fusion model provided greater clinical net benefit than any of the single-parameter models in both the internal and external validation cohorts (Figure 4E,4F). According to the DeLong test, in the training cohort, the differences between the fusion model and all single-parameter models were statistically significant (P<0.01). However, in the internal and external validation cohorts, the differences between the fusion model and the V-m-map and V-s-map models did not reach statistical significance (P>0.05) (see Figures S1,S2).

Figure 4 Model performance evaluation using ROC curves, calibration curves, and DCA. ADC, apparent diffusion coefficient; ANB, angiogenic branching index; AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic; V-m, mean flow velocity; V-s, velocity standard deviation.

Discussion

In recent years, deep learning models have demonstrated significant advantages in the field of medical research, surpassing the limitations of traditional qualitative or semi-quantitative imaging approaches in assessing pathological indicators and prognosis in patients with RC (17,18). In this study, a deep learning model was developed based on multi-b-value DWI-derived microvascular simulation parametric maps (V-m, V-s, and ANB) combined with the ADC-map, aiming to provide a preoperative non-invasive assessment of LVI status in RC patients. The results showed that, compared with single-parameter models, the fusion model achieved excellent performance, with AUCs of 0.901 and 0.863 in the internal and external validation cohorts, respectively, indicating a significant improvement in LVI discrimination capability. Despite the widespread use of MRI as the cornerstone imaging modality for preoperative RC assessment, real-world practice still exhibits considerable variability in imaging protocols and clinical workflows. This underscores the need for more standardized and clinically relevant tools to improve consistency and reliability in preoperative evaluation (19).

As a key indicator of tumor invasiveness in the progression of RC, LVI is highly dependent on abnormal microvascular formation, disordered perfusion, and hemodynamic disturbances. Previous studies have demonstrated that vascular remodeling within the tumor microenvironment not only facilitates tumor cell invasion and hematogenous dissemination but also markedly influences local perfusion status and hemodynamic characteristics, which are difficult to accurately capture and quantify on conventional T2WI or DWI images (20). In this study, a hemodynamic-diffusion coupled model was constructed by integrating the Navier-Stokes and Bloch-Torrey equations. Based on multi-b-value DWI signal inversion, three biophysically meaningful parametric maps were generated: V-m-map, representing local perfusion level; V-s-map, reflecting blood flow stability and variability; and ANB-map, indicating microvascular structural complexity and branching degree (21,22). Together, these parameters formed a multidimensional spatial representation of the tumor microcirculatory state, providing a structural-functional foundation for subsequent deep learning analysis. The Grad-CAM visualization further demonstrated that the model’s attention regions were concentrated along the tumor invasive front and in peritumoral areas with high vascular density, which were highly consistent with the typical spatial pathological distribution of LVI. These findings provide supportive visual evidence and suggest potential biological relevance (23,24).

In recent years, several studies have attempted to assess LVI in RC using MRI-based radiomics or deep learning approaches, with encouraging results. Tong et al. (25) developed a radiomics model based on dual-parameter MRI (T2WI and DWI), reporting an AUC of 0.95 for LVI assessment in RC. Fang et al. (26) combined DCE-MRI texture features with clinical factors and achieved an AUC of 0.877 in the validation cohort. Zhang et al. (27) constructed a multimodal radiomics model integrating CT and MRI, successfully assessing the LVI status in RC patients, with an AUC of 0.876 in the validation cohort. Shi et al. (28) also reported favorable performance using a multiparametric MRI-based deep learning model. Although the external validation AUC of the present model (0.863) was lower than some previously reported values, this difference may partly relate to variations in cohort composition, imaging protocols, and validation design across studies. In particular, the present study included an independent external validation cohort, whereas some previous studies were conducted in single-center settings. In medical imaging, performance decline during external validation is common because of differences in scanner hardware, acquisition protocols, and patient characteristics. Accordingly, the external validation results of the present study may provide a more realistic estimate of model generalizability (29). Despite these advancements, most previous studies have relied on conventional MRI sequences or radiomics-derived features, lacking a direct modeling of tumor microcirculatory hemodynamics. In contrast, the present study modeled the multi-b-value DWI signal from a biophysical perspective and constructed microvascular simulation parametric maps—namely, V-m, V-s, and ANB—which have physiologically relevant interpretations. These parameters provide a comprehensive representation of local blood flow distribution, perfusion variability, and vascular remodeling complexity (30). Furthermore, a ViT architecture was employed to model the microvascular simulation parametric maps. The multi-head self-attention mechanism of ViT enables the capture of long-range dependencies and global spatial structural information, making it well-suited to represent complex intratumoral blood flow patterns and microvascular structural characteristics (31). Building upon the ViT-extracted features, a Transformer encoder-based cross-parametric feature fusion module was further designed. Through multi-layer self-attention, this module achieved deep semantic fusion among different parametric maps. This structure enhanced the joint learning capability of multiparametric features and improved the integration of complementary information across parametric maps (32). The proposed fusion strategy demonstrated improved discriminative performance compared with single-parameter models in the external validation cohort, supporting the effectiveness of the structural design. Nevertheless, the present study primarily compared the fusion model with its single-parameter variants rather than with established clinical-radiological or multimodal baseline models. Therefore, the incremental value of the proposed approach should be interpreted with caution. Importantly, the primary contribution of the present study lies not only in performance improvement, but also in the introduction of a biologically informed framework that may provide complementary value beyond conventional signal-based approaches (33).

This study has several limitations. First, the calculation of microvascular simulation parameters depends on preset model assumptions (such as boundary conditions of blood flow and blood viscosity), making fully individualized modeling still challenging. Second, the regions of interest were manually delineated. Although double-blind evaluation and expert review were performed, subjective bias may still exist. Third, this study did not incorporate other MRI sequences such as T2WI or DCE-MRI; future work may explore multimodal fusion strategies to further improve model performance. Fourth, patients who received preoperative treatment were excluded from this study. As neoadjuvant therapy is commonly used in patients with locally advanced RC, this exclusion may limit the generalizability of the proposed model to such clinical scenarios. Fifth, some technical details of the microvascular simulation and deep learning model were not fully described in the original submission, which may affect reproducibility. Finally, although external validation was performed, the generalizability of the model across different scanners, imaging protocols, and clinical settings remains to be further investigated. In addition, as this was a retrospective study, further prospective, multicenter studies are warranted to validate the clinical applicability of the proposed model.

Conclusions

In conclusion, this study developed and validated a deep learning model integrating MRI-derived microvascular network simulation parameters for the preoperative assessment of LVI in RC. By incorporating biologically informed microvascular features, the proposed approach demonstrated encouraging performance in both internal and external validation cohorts. Grad-CAM visualization indicated that the model’s attention was mainly focused on tumor invasive fronts and peritumoral regions. These findings provide supportive visual evidence and suggest potential biological relevance. Overall, the proposed framework represents a promising and biologically motivated approach for noninvasive LVI assessment. However, further prospective and multicenter studies are required before clinical application.

Acknowledgments

The authors greatly appreciate all the patients and their families for participating in this trial. We also express our gratitude to the staffs members from The Affiliated Huai’an Hospital of Xuzhou Medical University and the No. 904 Hospital of the Joint Logistics Support Force of the PLA for their selfless dedication.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0226/rc

Data Sharing Statement: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0226/dss

Peer Review File: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0226/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0226/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of No. 904 Hospital of the Joint Logistics Support Force of the PLA (No. 20260116) and the Ethics Committee of The Affiliated Huai’an Hospital of Xuzhou Medical University (No. HEYLL20230053). Written informed consent was waived due to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Glynne-Jones R, Wyrwicz L, Tiret E, et al. Rectal cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol 2017;28:iv22-40. [Crossref] [PubMed]
2. Paul S, Arya S, Mokul S, et al. Extramural vascular invasion as an independent prognostic marker in locally advanced rectal cancer: propensity score match pair analysis. Abdom Radiol (NY) 2022;47:3671-8. [Crossref] [PubMed]
3. Djuric M, Kožik B, Vasiljevic T, et al. Prognostic Value of Separate Extramural Vascular Invasion Reporting in Operative Samples of Rectal Cancer: Single-Institutional Experience. Cancers (Basel) 2024;16:3579. [Crossref] [PubMed]
4. Maerten P, Defraene G, Wolthuis A, et al. Long-term recurrence patterns in rectal cancer after neoadjuvant therapy: 25-years of real-world data. Eur J Cancer 2026;234:116199. [Crossref] [PubMed]
5. Ammirati CA, Arezzo A, Gaetani C, et al. Can we apply the concept of sentinel lymph nodes in rectal cancer surgery. Minim Invasive Ther Allied Technol 2024;33:334-40. [Crossref] [PubMed]
6. Kožik B, Čorbo T, Pojskić N, et al. Identifying Molecular Modulators of the Vascular Invasion in Rectal Carcinoma: Role of ADAMTS8 and Its Co-Dependent Genes. Int J Mol Sci 2025;26:6261. [Crossref] [PubMed]
7. Dawson H, Bokhorst J, Studer L, et al. Lymph node metastases and recurrence in pT1 colorectal cancer: Prediction with the International Budding Consortium Score—A retrospective, multi-centric study. United European Gastroenterology Journal 2024;12:299-308.
8. Huang T, Zeng Y, Jiang R, et al. Interpretable MRI Subregional Radiomics-Deep Learning Model for Preoperative Lymphovascular Invasion Prediction in Rectal Cancer: A Dual-Center Study. J Imaging Inform Med 2026;39:1371-82. [Crossref] [PubMed]
9. Ning X, Yang D, Ao W, et al. A novel MRI-based radiomics for preoperative prediction of lymphovascular invasion in rectal cancer. Abdom Radiol (NY) 2025;50:3377-90. [Crossref] [PubMed]
10. Mayumi Takamune D, Miranda J, Mariussi M, et al. Diffusion-weighted imaging in rectal cancer MRI from theory to practice. Abdom Radiol (NY) 2026;51:578-98. [Crossref] [PubMed]
11. Luengo Gómez D, García Cerezo M, López Cornejo D, et al. The Value of MRI-Based Radiomics in Predicting the Pathological Nodal Status of Rectal Cancer: A Systematic Review and Meta-Analysis. Bioengineering (Basel) 2025;12:786. [Crossref] [PubMed]
12. Yazdi SAM, Miri S, Moghtadaie A, Nazar E. CD31 Positivity and Prediction of Tumor Invasive Behavior in Colorectal Cancer. International Journal of Cancer Management 2023;16.
13. van Valenberg W, Klein S, Vos FM, et al. An Efficient Method for Multi-Parameter Mapping in Quantitative MRI Using B-Spline Interpolation. IEEE Trans Med Imaging 2020;39:1681-9. [Crossref] [PubMed]
14. Caiazzo A, Heltai L, Vignon-Clementel IE. Mathematical modeling of blood flow in the cardiovascular system. In: Ambrosi D, Quarteroni A, Rozza G, editors. Modeling of Physiological Flows. Milan; New York: Springer; 2012: 39-61.
15. Kalamkar S, Amalanathan GM. MDA-ViT: Multimodal image fusion using dual attention vision transformer. Multimed Tools Appl 2025;84:23701-23.
16. Luo F, Wu D, Pino LR, et al. A novel multimodel medical image fusion framework with edge enhancement and cross-scale transformer. Sci Rep 2025;15:11657. [Crossref] [PubMed]
17. Ma Q, Meng R, Li R, et al. Multitask deep learning model based on multimodal data for predicting prognosis of rectal cancer: a multicenter retrospective study. BMC Med Inform Decis Mak 2025;25:209. [Crossref] [PubMed]
18. Jiang X, Zhao H, Saldanha OL, et al. An MRI Deep Learning Model Predicts Outcome in Rectal Cancer. Radiology 2023;307:e222223. [Crossref] [PubMed]
19. Boland PA, McEntee PD, Murphy E, et al. Local excision of rectal neoplasia: a real-world survey of current practices and perspectives. Minim Invasive Ther Allied Technol 2025;34:335-43. [Crossref] [PubMed]
20. Bates DDB, Homsi ME, Chang KJ, et al. MRI for Rectal Cancer: Staging, mrCRM, EMVI, Lymph Node Staging and Post-Treatment Response. Clin Colorectal Cancer 2022;21:10-8. [Crossref] [PubMed]
21. Lashgari M, Yang Z, Bernabeu MO, et al. SpinDoctor-IVIM: A virtual imaging framework for intravoxel incoherent motion MRI. Med Image Anal 2025;99:103369. [Crossref] [PubMed]
22. Piotrzkowska-Wróblewska H, Bajkowski JM, Dyniewicz B, et al. Identification of a spatially distributed diffusion model for simulation of temporal cellular growth. J Biomech 2025;182:112581. [Crossref] [PubMed]
23. Xiao B, Lv Y, Peng C, et al. Deep learning feature-based model for predicting lymphovascular invasion in urothelial carcinoma of bladder using CT images. Insights Imaging 2025;16:108. [Crossref] [PubMed]
24. Xu Y, Chen Y, Wu J, et al. The utility of Vision Transformer in preoperatively predicting microvascular invasion status of hepatocellular carcinoma. HPB (Oxford) 2023;25:533-42. [Crossref] [PubMed]
25. Tong P, Sun D, Chen G, et al. Biparametric magnetic resonance imaging-based radiomics features for prediction of lymphovascular invasion in rectal cancer. BMC Cancer 2023;23:61. [Crossref] [PubMed]
26. Fang J, Sun W, Wu D, et al. Value of texture analysis based on dynamic contrast-enhanced magnetic resonance imaging in preoperative assessment of extramural venous invasion in rectal cancer. Insights Imaging 2022;13:179. [Crossref] [PubMed]
27. Zhang Y, He K, Guo Y, et al. A Novel Multimodal Radiomics Model for Preoperative Prediction of Lymphovascular Invasion in Rectal Cancer. Front Oncol 2020;10:457. [Crossref] [PubMed]
28. Shi S, Singh A, Ma J, et al. Development and validation of a multi-parametric MRI deep-learning model for preoperative lymphovascular invasion evaluation in rectal cancer. Quant Imaging Med Surg 2025;15:427-39. [Crossref] [PubMed]
29. Brender JR, Ota M, Nguyen N, et al. Predicting ADC map quality from T2-weighted MRI: A deep learning approach for early quality assessment to assist point-of-care. Eur J Radiol 2025;191:112317. [Crossref] [PubMed]
30. Liu MM, Gladytz T, Dyke J, et al. Estimation of multicomponent flow in the kidney with multi-b-value spectral diffusion. Magn Reson Med 2025;94:2550-66. [Crossref] [PubMed]
31. Aburass S, Dorgham O, Al Shaqsi J, et al. Vision Transformers in Medical Imaging: a Comprehensive Review of Advancements and Applications Across Multiple Diseases. J Imaging Inform Med 2025;38:3928-71. [Crossref] [PubMed]
32. Yue Y, Li N, Zhang G, et al. A transformer-guided cross-modality adaptive feature fusion framework for esophageal gross tumor volume segmentation. Comput Methods Programs Biomed 2024;251:108216. [Crossref] [PubMed]
33. Benzoni I, Fricano M, Borali J, et al. Fluorescence-guided mesorectal nodes harvesting associated with local excision for early rectal cancer: technical notes. Minim Invasive Ther Allied Technol 2025;34:290-6. [Crossref] [PubMed]