Deep learning model for predicting lymph node metastasis around rectal cancer based on rectal tumor core area and mesangial imaging features

Abstract

Assessing lymph node metastasis (LNM) involvement in patients with rectal cancer (RC) is fundamental in disease management. In this study, we used artificial intelligence (AI) technology to develop a segmentation model that automatically segments the tumor core area and mesangial tissue from magnetic resonance T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) images collected from 122 RC patients to improve the accuracy of LNM prediction, after which omics machine modeling was performed on the segmented ROI. An automatic segmentation model was developed using nn-UNet. This pipeline integrates deep learning (DL), specifically 3D U-Net, for semantic segmentation and image processing techniques such as resampling, normalization, connected component analysis, image registration, and radiomics features coupled with machine learning. The results showed that the DL segmentation method could effectively segment the tumor and mesangial areas from MR sequences (the median dice coefficient: 0.90 ± 0.08; mesorectum segmentation: 0.85 ± 0.36), and the radiological characteristics of rectal and mesangial tissues in T2WI and ADC images could help distinguish RC treatments. The nn-UNet model demonstrated promising preliminary results, achieving the highest area under the curve (AUC) values in various scenarios. In the evaluation encompassing both tumor lesions and mesorectum involvement, the model exhibited an AUC of 0.743, highlighting its strong discriminatory ability to predict a combined outcome involving both elements. Specifically targeting tumor lesions, the model achieved an AUC of 0.731, emphasizing its effectiveness in distinguishing between positive and negative cases of tumor lesions. In assessing the prediction of mesorectum involvement, the model displayed moderate predictive utility with an AUC of 0.753. The nn-UNet model demonstrated impressive performance across all evaluated scenarios, including combined tumor lesions and mesorectum involvement, tumor lesions alone, and mesorectum involvement alone.

Clinical trial number

Not applicable.

Introduction

Rectal cancer (RC) is the third most commo cancer and the second leading cause of cancer-related deaths worldwide [1]. It predominantly affects older individuals. Surgery remains the most effective treatment for those diagnosed with early-stage carcinoma. Nevertheless, considering that most patients present with late-stage cancer at the time of diagnosis [2], more invasive therapy is often required. Thus, searching for accurate prediction tools is essential.

Lymph node metastasis (LNM) is the main mode of metastasis that occurs in 30–40% of patients with RC [3]. Thus, accurate LNM prediction is crucial for determining RC prognosis [4]. Positron emission tomography/computed tomography (PET/CT) [5], magnetic resonance imaging (MRI) [6], and endoscopic ultrasonography (EUS) [7] are considered the best non-invasive tools for assessing the LNM of RC. However, EUS does not appear helpful in post-therapeutic response assessments [8]. On the other hand, the utility of CT in identifying and following up metastatic lesions is limited [9]. MRI, the most widely used tool, can detect precise rectum and mesenteric fascia anatomy and accurately distinguish benign from malignant lymph nodes (LN) [10]. Also, MRI predicts the circumferential resection margin and tumor staging [11]. However, for early-stage carcinoma, the utility of the boundary value of LN diameter remains inconclusive, and different studies have encountered the problems of under and over-staging [12].

Over the years, conventional radiomics have been developed and applied to analyze RC lesions on MR images [13,14,15]. However, these methods often require manual delineation of the region of interest (ROIs); manual drawing of ROIs onto the tumor for quantitative or qualitative assessment is time-consuming and may result in interobserver variation. Thus, artificial intelligence (AI) methods, such as machine learning and DL tools, have been increasingly used to improve LNM assessment in patients with RC [16]. These AI methods can automatically extract regions and tissues of interest from original MR images. Compared to traditional imaging, AI models can obtain rapid results using predictive algorithms, high accuracy, adaptable models, and standardized results across different systems. Following the application of AI technology in diagnosing LNM, the clinical outcomes and prognosis of RC have improved significantly in China [17].

Previous studies reporting on AI models have mainly relied on tumor core area to predict LNM [17, 18], while few have considered the mesentery. Luo et al. [19] found metastatic cancer cells in the mesentery of colorectal cancer patients and suggested that these kinds of clinical features are associated with poor prognosis. In this study, we used AI technology to develop a segmentation model that automatically segmented the tumor core area and mesangial tissue from magnetic resonance T2WI and ADC images collected from 122 preoperative RC patients to improve the accuracy of LNM prediction. Our entire AI process was divided into two stages: images were automatically segmented, after which the omics machine modeling was performed on the segmented ROI. These findings underscore the potential of the developed AI model to significantly impact clinical practice by enhancing the identification of LNM, ultimately contributing to improved patient care and survival in RC treatment.

Materials and methods

Patients

This study followed the Declaration of Helsinki and obtained approval from the Second Affiliated Hospital Ethics Committees, Harbin Medical University. The ethical approval number is YJSKY2023-301. The requirement for written informed consent was waived.

A total of 261 patients diagnosed with RC who underwent MRI at the Second Affiliated Hospital, Harbin Medical University, between December 2020 and December 2022 were enrolled in this retrospective study. The inclusion criteria were: (1) confirmation of RC with LNM through postoperative pathological examination based on Chinese Society of Clinical Oncology (CSCO) criteria [20]; (2) presence of a single lesion; (3) baseline rectal MRI examination conducted within14 days prior to surgical resection. Exclusion criteria were the following: (1) poor image quality (n = 33); (2) receipt of any systemic or local treatment before surgical resection, such as neoadjuvant chemoradiotherapy (n = 28); (3) patients with other malignancies (n = 12); (4) palliative resection (n = 57) or history of previous pelvic surgery (n = 9). The flowchart is shown in Fig. 1.

[IMAGE OMITTED: SEE PDF]

Image acquisition

MRI scans were performed using a 3.0 T whole-body scanner (GE Discovery 750 W, WI). In order to ensure optimal image quality, patients consumed liquid food the day before the examination and were administered 20 mL glycerin enema to ensure that the rectum was clear and empty 3 to 4 h before the examination. High-resolution rectal MRI protocols included sagittal fat suppression, transverse T2WI, and transverse T1WI. The acquisition parameters included: flip angle, 110°; repetition time (TR)/echo time (TE), 5373/96 ms; field of view (FOV), 26 cm2; slices, 24; matrix size, 256 × 256; slice thickness, 5 mm; spacing between slices, 1 mm. ADC maps were generated by the post-processing on the MR system.

Pathological characteristics

Pathological reports of surgically resected specimens were collected and assessed using histopathologic assessment [21]; the absence of regional LNM was recognized as negative, while the LNM was defined as positive when the number of regional LN metastasis was ≥ 1.

Tumor segmentation and radiomics analysis

Our entire AI process was divided into two stages: images were automatically segmented and omics machine modeling was performed on the segmented ROI (Fig. 2).

[IMAGE OMITTED: SEE PDF]

Tumor segmentation

First, images were automatically segmented. Briefly, 122 cases were included in the final analysis; 85 patients were included in the training set and 37 in the test set, according to a 7:3 ratio. Two experienced radiologists, each with 12 and 7 years of expertise in RC diagnoses, conducted manual segmentation of RC lesions and perirectal tissues on T2-weighted imaging (T2WI) using ITK-snap (v3.8.0, http://www.itksnap.org). Patient information was intentionally concealed during this process. Following the definition from previous studies [18], peri-tumoral regions were defined as the mesorectal fat surrounding the tumor, extending from the mesorectal fascia to the edge of the tumor.

Automatic segmentation was performed using the nn-UNet [23] on T2 images. The nn-UNet, leveraging the self-configuring capabilities of U-Net, demonstrated consistent excellence in DL-based biomedical image segmentation. To enhance the segmentation process, pre-processing techniques, including image resampling and normalization, were applied to the T2 images before feeding them into the 3D Unet network for training. Post-processing involved connected component analysis. The output of this stage included image registration using SimpleElastix, with T2 images serving as inputs and DWI and ADC images undergoing rigid registration to the segmented T2 results (Fig. 2). Furthermore, nn-UNet (Fig. 3) architecture involved non-zero cropping, normalization, and a 3D-Unet structure with convolutional downsampling, convolution, bilinear upsampling, and a dense feature stack. The training and inference processes incorporated five-fold cross-validation, data augmentation with a 50% overlap, and post-processing involving connected component analysis. The final output comprised segmented regions of interest (ROIs) for ADC, T1, and T2, along with radiomics features. Further analysis included feature selection and the application of machine learning models to assess the lymph node metastasis (LNM) status.

[IMAGE OMITTED: SEE PDF]

Feature extraction and radiomics analysis

The integrative pipeline for radiomics analysis in LNM prediction using 3D Unet and image registration is shown in Fig. 2. In the feature extraction process, the segmented regions, specifically the ADC ROI, T1 ROI, and T2 ROI, obtained through neural network segmentation have a crucial role. PyRadiomics (available at http://github.com/radiomics/pyradiomics) was employed to extract relevant features from these segmented regions. RadiomicsFeatureExtractor instance was initialized, and extraction settings, including filters and discretization parameters, were configured. Subsequently, the initialized extractor was applied to compute radiomic features independently for each segmented region. The shape, first-order, glcm, glrlm, glszm and gldm features were extracted. Then we used Variance(threshold = 1) and LASSO regularization to select features.

The resulting feature vectors obtained from the ADC, T1, and T2 modalities were concatenated, forming a comprehensive feature vector. Optionally, feature selection techniques can be incorporated into the process to address dimensionality concerns. These techniques refine the feature vector, ensuring that only the most relevant and impactful features are retained for further analysis.

This concatenated and, if applicable, selected feature vector serves as the input for the subsequent phase, which involves the construction of a machine-learning model. The primary objective of this model was to predict the LNM status based on the extracted radiomic information. It is worth noting that adjustments to settings and parameters can be made during this process to accommodate specific dataset characteristics and meet analysis requirements effectively.

The constructed machine learning model was evaluated using the AUC, a widely recognized metric in classification tasks. The AUC comprehensively assesses the ability of the model to discriminate between positive and negative instances. This evaluation process ensures the robustness and effectiveness of the model in predicting LNM.

As a crucial step in the model refinement process, a key feature is selected based on its significance and contribution to the predictive performance of the model. This selected key feature is pivotal in finalizing the machine learning model, enhancing its interpretability, and facilitating a more targeted and focused understanding of the underlying radiomic information relevant to Lymph Node Metastasis prediction.

Statistical analysis

Software tools (MedCalc software version 11.2, Python version 3.5, and SPSS version 24.0) were used for statistical analysis. Normality testing of all continuous variables was conducted using the Kolmogorov–Smirnov test to assess their distribution. Categorical data were compared using a chi-square test, depending on the expected cell counts. For continuous variables, presented as mean ± standard deviation, comparisons were made using either the Student’s t-test for normally distributed data or the Kruskal–Wallis H test for variables with non-normal distributions. We used the Dice coefficient to evaluate the performance of the deep learning segmentation model.

Results

Table 1 summarizes the demographic and clinical characteristics of a group of patients. Among the participants, 67 (54.9%) were males and 55 (45.1%) were females. The average age of the patients was 58.4 years, with a standard deviation of 9.1 years, indicating the central tendency and dispersion of age within the group.

[IMAGE OMITTED: SEE PDF]

Automatic segmentation results

A total of 85 patients were included in the training set and 37 in the test set according to a 7:3 ratio. These automatic segmentation models exhibited promising preliminary results in the test datasets, as illustrated in Fig. 4; Table 2.

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

Figure 4 presents a visual comparison of lesion and mesorectum segmentation results. The segmentation analysis was performed on T2-weighted imaging (T2WI) to assess the accuracy and performance of the AI model in comparison to human expertise. Panel A shows the original T2WI image, Panel B represents the AI model prediction, and Panel C displays the segmentation by an experienced radiologist.

Color-coded overlays and contour lines are used to visually represent the segmented regions, facilitating a clear comparison between the predictions of the AI model and the radiologist’s manual annotations. This figure provides a comprehensive assessment of the segmentation accuracy of the AI model across different imaging modalities, contributing valuable insights for the integration of AI in clinical radiology workflows.

Table 2 compares segmentation accuracy in different regions, specifically the tumor and mesorectum, using the median dice coefficient. The dice coefficient is a metric commonly used to assess the similarity between two segmentation masks, with values ranging from 0 (no overlap) to 1 (perfect overlap) [22]. For tumor segmentation, Reader 1 vs. Reader 2 demonstrated a median dice coefficient of 0.93 ± 0.16, indicating a high degree of overlap between the segmentations by these two readers. Similarly, for mesorectum segmentation, the comparison between Reader 1 and 2 yielded a median dice coefficient of 0.87 ± 0.26. The comparisons between Reader 1 and nnU-Net showed slightly lower median dice coefficients, indicating differences in segmentation accuracy between the human reader and the automated neural network model. Specifically, the median dice coefficient for tumor segmentation was 0.90 ± 0.08, and for mesorectum segmentation, it was 0.85 ± 0.36. These values provide insights into the performance of different segmentation methods in accurately delineating the specified regions.

Performance of different models in the training set and validation set

As shown in Table 3. the logistic regression (LR) model demonstrated good discriminative ability with an AUC of 0.846 and 0.853 in the training and validation set, respectively. Accuracy (ACC) was moderate, with values of 0.672 (training) and 0.660 (validation). The sensitivity and specificity values indicate the ability of the model to correctly identify positive and negative cases, respectively. Positive Predictive Value (PPV) and Negative Predictive Value (NPV) were also reported, providing insights into the performance of the model in predicting true positives and true negatives.

[IMAGE OMITTED: SEE PDF]

The random forest (RF) model showed a decent AUC of 0.800 in the training set but a lower AUC of 0.711 in the validation set. ACC was consistent in both sets at 0.672 (training) and 0.623 (validation). Sensitivity was relatively low in the validation set, indicating challenges in identifying true positives, while specificity was higher in the validation set, suggesting better performance in correctly identifying true negatives.

Support Vector Machine (SVM) and Decision Tree (DT) models exhibited competitive AUC values and ACC in the training and validation sets. SVM demonstrates good specificity, while DT showed higher sensitivity.

Bayesian Model (Bayes) model had a moderate AUC and ACC, with sensitivity and specificity values indicating a balanced performance in identifying positive and negative cases. K-Nearest Neighbors (KNN) showed good discriminative ability with a high AUC in the training set but a lower AUC in the validation set. ACC was consistent, and the model exhibits balanced sensitivity and specificity. Gradient Boosting Decision Tree (GBDT) had a lower AUC and ACC, with challenges in sensitivity and PPV, indicating limitations in correctly identifying positive cases.

Radiologists 1 and 2 achieved high ACC and good sensitivity and specificity, indicating strong performance in classifying positive and negative cases.

Discussion

Numerous research studies have highlighted the potential of AI to revolutionize healthcare, particularly in applications involving image recognition [23]. AI has been employed for assessing colonic polyps [24], adenomas [25], colorectal cancer [26], ulcerative colitis [27], and intestinal motility disorders in colorectal diseases [28]. The rapid advancement of AI technology ensures its continuous and pivotal role in colorectal diagnosis and treatment [29]. With increasing computing power and access to extensive imaging databases, there is an opportunity to develop more accurate AI algorithms. Deep learning (DL) applications for medical imaging are currently popular; however, these models have drawbacks such as sensitivity to image variability, large sample size requirements, poor generalization, and extensive computing resource needs [30]. Also, DL models may struggle with distribution shifts caused by diverse imaging acquisition parameters and scanner types [31]. In this study, we developed a fully automated segmentation model based on nn-UNet to precisely segment RC from T2W and ADC images. Subsequently, we constructed DL-based classification models that significantly improved performance in delineating lesions. Integrating advanced radiomics and DL techniques, particularly the nn-UNet model, has demonstrated promising preliminary results in accurately segmenting tumor lesions and perirectal tissues from T2WI and ADC maps obtained through high-resolution rectal MRI scans. These findings underscore the potential clinical impact of the developed AI model in enhancing the identification of LNM, ultimately contributing to improved patient care and survival in RC treatment.

The presence of cancer cells in the regional lymph node is an early sign of metastasis in patients with colon cancer [18]. In 2021, Bedrikovetski et al. [32] performed a meta-analysis to evaluate the diagnostic accuracy of AI models that can detect LNM on pre-operative staging imaging for colorectal cancer. However, the mesentery was not considered by any of these studies [32]. Hacim et al. [33] found that the mesenteric area is significantly correlated with total LN counts. Luo and colleagues [19] further discovered metastatic cancer cells in the mesentery of colorectal cancer patients and suggested that these kinds of clinical features are associated with poor prognosis. The study concludes by emphasizing the significance of combining radiomic parameters of tumor lesions and perirectal tissues, as evidenced by the AUC results in Table 3. This combination marginally enhances the accuracy of predicting postoperative LNM in patients with RC, underscoring the potential clinical utility of the developed automated segmentation and classification models.

Most studies reporting AI models for detecting LNM on pre-operative staging imaging for colorectal cancer utilized radiomics rather than the more recent deep learning methodology, largely as the latter has the most remarkable novelty and the need for specialized expertise [34,35,36]. This disproportion in representation hinders definitive comparisons between the two AI models. Additionally, these studies had a retrospective design, rendering them susceptible to confounding and selection bias. Also, technical aspects of algorithms were emphasized in several studies, while key limitations, including input variation, lack of clinical information, and potential data overfitting, were often overlooked. The model we introduced in the present study is a fully automated method that combines the advantages of the above two types of research. Leveraging advanced radiomics and DL techniques, this new method integrates state-of-the-art technologies to enhance the accuracy and efficiency of pre-operative assessment. High-resolution rectal MRI scans, encompassing T2WI and ADC maps, allow for detailed characterization of tumor lesions and perirectal tissues. A distinctive aspect of our methodology lies in integrating a 3D Unet neural network for automated segmentation, thus providing precise delineation of ROI. This process is complemented by a radiomics feature extraction pipeline, utilizing PyRadiomics for robust feature computation. The resulting comprehensive feature vector, derived from ADC, T1, and T2 modalities, serves as input for machine learning models, facilitating the prediction of LNM status. The seamless integration of automated segmentation, radiomics, and machine learning underscores the novelty of our approach, offering a holistic and data-driven solution for accurate prediction of LNM in RC patients. However, clinical utility claims, such as the potential to reduce overtreatment by enabling more accurate preoperative staging and decision-making, require prospective validation. Future studies should explore perirectal tissue imaging features to improve prediction accuracy and clinical application in the management of RC patients.

Limitations

Despite the promising outcomes of our proposed methodology, several limitations must be acknowledged. Firstly, the retrospective nature of the study introduces inherent biases and limits the establishment of causality. The reliance on a single-center dataset may also impact the generalizability of the findings. Additionally, while efforts were made to ensure rigorous inclusion and exclusion criteria, the possibility of selection bias cannot be completely ruled out. The exclusion of patients who received neoadjuvant chemoradiotherapy or had a history of previous pelvic surgery may limit the applicability of our findings to a broader patient population. Furthermore, although demonstrating high performance, the reliance on a 3D Unet architecture for automated segmentation may be sensitive to variations in imaging protocols and equipment. As with any predictive modeling, the potential for overfitting and the need for external validation should be considered. Lastly, the absence of prospective validation and real-time clinical application warrants caution in translating our method into routine clinical practice. These limitations underscore the need for future multi-center, prospective studies to validate and refine the proposed methodology for enhanced clinical utility and generalizability.

Conclusion

This study introduces a novel and comprehensive approach to predicting postoperative LNM in RC patients. The automated segmentation model consistently exhibited impressive performance across various scenarios, including predicting combined tumor lesions and mesorectum involvement, tumor lesions alone, and mesorectum involvement alone. The high AUC values achieved, especially in scenarios involving both elements, emphasize the strong discriminatory ability of the model in predicting a combined outcome. The study also highlights the advantages of combining radiomic parameters of tumor lesions and perirectal tissues, showcasing a marginal enhancement in predicting postoperative LNM.

Data availability

The datasets used in this study are not publicly available due to privacy.Correspondence: Kuang Fu Email: [email protected].

Abbreviations

LNM:

Lymph node metastasis

RC:

Rectal cancer

AI:

Artificial intelligence

T2WI:

Magnetic resonance T2-weighted imaging

ADC:

Apparent diffusion coefficient

PET/CT:

Positron emission tomography/computed tomography

MRI:

Magnetic resonance imaging

EUS:

Endoscopic ultrasonography

LN:

Malignant lymph nodes

ROIs:

Region of interest

LR:

Logistic regression

RF:

Random forest)

SVM:

Support vector machine

DT:

Decision tree

KNN:

K-nearest neighbors

GBDT:

Gradient boosting decision tree

References

Rawla P, Sunkara T, Barsouk A. Epidemiology of colorectal cancer: incidence, mortality, survival, and risk factors. Prz Gastroenterol. 2019;14(2):89–103. https://doi.org/10.5114/pg.2018.81072. Epub 20190106. PubMed PMID: 31616522; PubMed Central PMCID: PMCPMC6791134.

Shen D, Wang P, Xie Y, Zhuang Z, Zhu M, Wang X, et al. Clinical spectrum of rectal cancer identifies hallmarks of early-onset patients and next-generation treatment strategies. Cancer Med. 2023;12(3):3433–41. https://doi.org/10.1002/cam4.5120. PubMed PMID: 35929660; PubMed Central PMCID: PMCPMC9939204.

Jin M, Frankel WL. Lymph node metastasis in colorectal cancer. Surg Oncol Clin N Am. 2018;27(2):401–12. PubMed PMID: 29496097.

Peng W, Qiao H, Mo L, Guo Y. Progress in the diagnosis of lymph node metastasis in rectal cancer: a review. Front Oncol. 2023;13:1167289. Epub 20230713. https://doi.org/10.3389/fonc.2023.1167289. PubMed PMID: 37519802; PubMed Central PMCID: PMCPMC10374255.

Dahmarde H, Parooie F, Salarzaei M. Is (18)F-FDG PET/CT an accurate way to detect lymph node metastasis in colorectal cancer: a systematic review and meta-analysis. Contrast Media Mol Imaging. 2020;2020:5439378. Epub 20200718. https://doi.org/10.1155/2020/5439378. PubMed PMID: 32733174; PubMed Central PMCID: PMCPMC7383332.

Beets-Tan RGH, Lambregts DMJ, Maas M, Bipat S, Barbaro B, Curvo-Semedo L, et al. Magnetic resonance imaging for clinical management of rectal cancer: updated recommendations from the 2016 European society of Gastrointestinal and abdominal radiology (ESGAR) consensus meeting. Eur Radiol. 2018;28(4):1465–75. https://doi.org/10.1007/s00330-017-5026-2. Epub 20171017.

Yan L, Weifeng Z, Qin W, Jinping W. A model based on endorectal ultrasonography predicts lateral lymph node metastasis in low and middle rectal cancer. J Clin Ultrasound. 2022;50(5):705–12. https://doi.org/10.1002/jcu.23204. Epub 20220324.

Krajewski KM, Kane RA. Ultrasound staging of rectal cancer. Semin Ultrasound CT MR. 2008;29(6):427–32. https://doi.org/10.1053/j.sult.2008.10.005. PubMed PMID: 19166040.

Heo SH, Kim JW, Shin SS, Jeong YY, Kang HK. Multimodal imaging evaluation in staging of rectal cancer. World J Gastroenterol. 2014;20(15):4244–55. https://doi.org/10.3748/wjg.v20.i15.4244. PubMed PMID: 24764662; PubMed Central PMCID: PMCPMC3989960.

Zhuang Z, Zhang Y, Wei M, Yang X, Wang Z. Magnetic resonance imaging evaluation of the accuracy of various lymph node staging criteria in rectal cancer: a systematic review and meta-analysis. Front Oncol. 2021;11:709070. PubMed PMID: 34327144; PubMed Central PMCID: PMCPMC8315047.

Grone J, Loch FN, Taupitz M, Schmidt C, Kreis ME. Accuracy of various lymph node staging criteria in rectal cancer with magnetic resonance imaging. J Gastrointest Surg. 2018;22(1):146 – 53. https://doi.org/10.1007/s11605-017-3568-x. Epub 20170912. PubMed PMID: 28900855.

Nougaret S, Reinhold C, Mikhael HW, Rouanet P, Bibeau F, Brown G. The use of MR imaging in treatment planning for patients with rectal carcinoma: have you checked the DISTANCE? Radiology. 2013;268(2):330–44. https://doi.org/10.1148/radiol.13121361. PubMed PMID: 23882096.

Li H, Chen XL, Liu H, Lu T, Li ZL. MRI-based multiregional radiomics for predicting lymph nodes status and prognosis in patients with resectable rectal cancer. Front Oncol. 2022;12:1087882. https://doi.org/10.3389/fonc.2022.1087882. Epub 20230104.

Niu Y, Yu X, Wen L, Bi F, Jian L, Liu S, et al. Comparison of preoperative CT- and MRI-based multiparametric radiomics in the prediction of lymph node metastasis in rectal cancer. Front Oncol. 2023;13:1230698. PubMed PMID: 38074652; PubMed Central PMCID: PMCPMC10708912.

Qin S, Lu S, Liu K, Zhou Y, Wang Q, Chen Y et al. Radiomics from mesorectal blood vessels and lymph nodes: a novel prognostic predictor for rectal cancer with neoadjuvant therapy. Diagnostics (Basel). 2023;13(12). Epub 20230606. https://doi.org/10.3390/diagnostics13121987. PubMed PMID: 37370882; PubMed Central PMCID: PMCPMC10297215.

Wang PP, Deng CL, Wu B. Magnetic resonance imaging-based artificial intelligence model in rectal cancer. World J Gastroenterol. 2021;27(18):2122–30. https://doi.org/10.3748/wjg.v27.i18.2122. PubMed PMID: 34025068; PubMed Central PMCID: PMCPMC8117733.

Gao Y, Zhang XX, Li S, Lu Y. [Application of artificial intelligence technology in the diagnosis and treatment of colorectal cancer]. Zhonghua Wei Chang Wai Ke Za Zhi. 2020;23(12):1155–8. https://doi.org/10.3760/cma.j.cn.441530-20200213-00054. PubMed PMID: 33353269.

Ong ML, Schofield JB. Assessment of lymph node involvement in colorectal cancer. World J Gastrointest Surg. 2016;8(3):179–92. https://doi.org/10.4240/wjgs.v8.i3.179. PubMed PMID: 27022445; PubMed Central PMCID: PMCPMC4807319.

Luo XL, Xie DX, Wu JX, Wu AD, Ge ZQ, Li HJ, et al. Detection of metastatic cancer cells in mesentery of colorectal cancer patients. World J Gastroenterol. 2017;23(34):6315–20. https://doi.org/10.3748/wjg.v23.i34.6315. PubMed PMID: 28974898; PubMed Central PMCID: PMCPMC5603498.

Diagnosis TG, For Colorectal Cancer Working Group C. Chinese society of clinical oncology (CSCO) diagnosis and treatment guidelines for colorectal cancer 2018 (English version). Chin J Cancer Res. 2019;31(1):117–34.https://doi.org/10.21147/j.issn.1000-9604.2019.01.07. PubMed PMID: 30996570; PubMed Central PMCID: PMCPMC6433585.

Oh JR, Park B, Lee S, Han KS, Youk EG, Lee DH, et al. Nomogram development and external validation for predicting the risk of lymph node metastasis in T1 colorectal cancer. Cancer Res Treat. 2019;51(4):1275–84. Epub 20190117. https://doi.org/10.4143/crt.2018.569.

Vali M, Nazari B, Sadri S, Pour EK, Riazi-Esfahani H, Faghihi H et al. CNV-Net: Segmentation, classification and activity score measurement of choroidal neovascularization (CNV) using optical coherence tomography angiography (OCTA). Diagnostics (Basel). 2023;13(7). Epub 20230331. https://doi.org/10.3390/diagnostics13071309. PubMed PMID: 37046527; PubMed Central PMCID: PMCPMC10093691.

Davenport T, Kalakota R. The potential for artificial intelligence in healthcare. Future Healthc J. 2019;6(2):94–8. https://doi.org/10.7861/futurehosp.6-2-94. PubMed PMID: 31363513; PubMed Central PMCID: PMCPMC6616181.

Li MD, Huang ZR, Shan QY, Chen SL, Zhang N, Hu HT, et al. Performance and comparison of artificial intelligence and human experts in the detection and classification of colonic polyps. BMC Gastroenterol. 2022;22(1):517. Epub 20221213. https://doi.org/10.1186/s12876-022-02605-2. PubMed PMID: 36513975; PubMed Central PMCID: PMCPMC9749329.

Adiwinata R, Tandarto K, Arifputra J, Waleleng BJ, Gosal F, Rotty L, et al. The impact of artificial intelligence in improving polyp and adenoma detection rate during colonoscopy: Systematic-Review and Meta-Analysis. Asian Pac J Cancer Prev. 2023;24(11):3655–63. PubMed PMID: 38019222; PubMed Central PMCID: PMCPMC10772777.

Kavitha MS, Gangadaran P, Jackson A, Venmathi Maran BA, Kurita T, Ahn BC. Deep neural network models for colon cancer screening. Cancers (Basel). 2022;14(15). Epub 20220729. https://doi.org/10.3390/cancers14153707. PubMed PMID: 35954370; PubMed Central PMCID: PMCPMC9367621.

Gubatan J, Levitte S, Patel A, Balabanis T, Wei MT, Sinha SR. Artificial intelligence applications in inflammatory bowel disease: emerging technologies and future directions. World J Gastroenterol. 2021;27(17):1920–35. https://doi.org/10.3748/wjg.v27.i17.1920. PubMed PMID: 34007130; PubMed Central PMCID: PMCPMC8108036.

Zulqarnain F, Rhoads SF, Syed S. Machine and deep learning in inflammatory bowel disease. Curr Opin Gastroenterol. 2023;39(4):294–300. Epub 20230508. https://doi.org/10.1097/MOG.0000000000000945. PubMed PMID: 37144491; PubMed Central PMCID: PMCPMC10256313.

Patel V, Khan MN, Shrivastava A, Sadiq K, Ali SA, Moore SR, et al. Artificial intelligence applied to Gastrointestinal diagnostics: A review. J Pediatr Gastroenterol Nutr. 2020;70(1):4–11. https://doi.org/10.1097/MPG.0000000000002507. PubMed PMID: 31567886; PubMed Central PMCID: PMCPMC6934912.

Li M, Jiang Y, Zhang Y, Zhu H. Medical image analysis using deep learning algorithms. Front Public Health. 2023;11:1273253. Epub 20231107. https://doi.org/10.3389/fpubh.2023.1273253. PubMed PMID: 38026291; PubMed Central PMCID: PMCPMC10662291.

Svanera M, Savardi M, Signoroni A, Benini S, Muckli L. Fighting the scanner effect in brain MRI segmentation with a progressive level-of-detail network trained on multi-site data. Med Image Anal. 2024;93:103090. Epub 20240117. https://doi.org/10.1016/j.media.2024.103090. PubMed PMID: 38241763.

Bedrikovetski S, Dudi-Venkata NN, Kroon HM, Seow W, Vather R, Carneiro G, et al. Artificial intelligence for pre-operative lymph node staging in colorectal cancer: a systematic review and meta-analysis. BMC Cancer. 2021;21(1):1058. Epub 20210926. https://doi.org/10.1186/s12885-021-08773-w. PubMed PMID: 34565338; PubMed Central PMCID: PMCPMC8474828.

Hacim NA, Akbas A, Ulgen Y, Aktokmakyan TV, Meric S, Tokocin M, et al. Influence of colonic mesenteric area on the number of lymph node retrieval for colon cancer: a prospective cohort study. Ann Coloproctol. 2023;39(1):77–84. Epub 20210916. https://doi.org/10.3393/ac.2021.00444.0063. PubMed PMID: 34525506; PubMed Central PMCID: PMCPMC10009066.

Yang YS, Feng F, Qiu YJ, Zheng GH, Ge YQ, Wang YT. High-resolution MRI-based radiomics analysis to predict lymph node metastasis and tumor deposits respectively in rectal cancer. Abdom Radiol (NY). 2021;46(3):873 – 84. Epub 20200917. https://doi.org/10.1007/s00261-020-02733-x. PubMed PMID: 32940755.

Nakanishi R, Akiyoshi T, Toda S, Murakami Y, Taguchi S, Oba K et al. Radiomics approach outperforms diameter criteria for predicting pathological lateral lymph node metastasis after neoadjuvant (chemo)radiotherapy in advanced low rectal cancer. Ann Surg Oncol. 2020;27(11):4273-83. Epub 20200807. https://doi.org/10.1245/s10434-020-08974-w. PubMed PMID: 32767224.

Zhou X, Yi Y, Liu Z, Zhou Z, Lai B, Sun K, et al. Radiomics-Based preoperative prediction of lymph node status following neoadjuvant therapy in locally advanced rectal cancer. Front Oncol. 2020;10:604. Epub 20200511. https://doi.org/10.3389/fonc.2020.00604.