Preoperative prediction of rectal Cancer staging combining MRI deep transfer learning, radiomics features, and clinical factors: accurate differentiation from stage T2 to T3

Abstract

Background

This study evaluates the efficacy of integrating MRI deep transfer learning, radiomic signatures, and clinical variables to accurately preoperatively differentiate between stage T2 and T3 rectal cancer.

Methods

We included 361 patients with pathologically confirmed stage T2 or T3 rectal cancer, divided into a training set (252 patients) and a test set (109 patients) at a 7:3 ratio. The study utilized features derived from deep transfer learning and radiomics, with Spearman rank correlation and the Least Absolute Shrinkage and Selection Operator (LASSO) regression techniques to reduce feature redundancy. Predictive models were developed using Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM), selecting the best-performing model for a comprehensive predictive framework incorporating clinical data.

Results

After removing redundant features, 24 key features were identified. In the training set, the area under the curve (AUC)values for LR, RF, DT, and SVM were 0.867, 0.834, 0.900, and 0.944, respectively; in the test set, they were 0.847, 0.803, 0.842, and 0.910, respectively. The combined model, using SVM and clinical variables, achieved AUCs of 0.946 in the trainingset and 0.920 in the test set.

Conclusion

The study confirms the utility of a combined model of MRI deep transfer learning, radiomic features, and clinical factors for preoperative classification of stage T2 vs. T3 rectal cancer, offering significant technological support for precise diagnosis and potential clinical application.

Introduction

Rectal cancer ranks among the predominant malignancies globally, witnessing a consistent rise in incidence over recent years [1–3], thus emerging as a significant contributor to cancer-related mortality [4]. The therapeutic modalities for rectal cancer encompass chemotherapy, radiotherapy, and surgical interventions, with the T stage serving as a critical determinant in the formulation of treatment strategies. According to the National Comprehensive Cancer Network (NCCN) guidelines, patients diagnosed with stage T2 rectal cancer, in the absence of lymph node or distant metastasis, are typically recommended to undergo total mesorectal excision (TME); conversely, stage T3 rectal cancer patients are advised to receive neoadjuvant chemoradiotherapy (nCRT) irrespective of lymph node or distant metastasis status [5, 6]. Consequently, precise preoperative differentiation between stages T2 and T3 of rectal cancer is imperative for the selection of an optimal treatment plan and for prognostic assessment.

While conventional MRI is indispensable for rectal cancer staging [7, 8], its efficacy in discriminating between stages T2 and T3 can be compromised by fibrotic reactions surrounding the tumor. The advent of artificial intelligence, especially the expanded application of machine learning and deep learning within the realm of medical imaging [9–12], has unveiled remarkable capabilities in molecular classification, prognosis forecasting, and evaluating therapeutic outcomes in rectal cancer. For instance, BILAL et al. [13] utilized the ResNet-34 deep learning architecture to prognosticate the genetic status of rectal cancer, attaining AUC metrics of 0.81, 0.86, and 0.83 for high mutation load, microsatellite instability, and chromosomal instability, respectively. YU et al. [14] identified age, gender, marital status, tumor grade, surgical history, and chemotherapy as independent predictors of survival in rectal cancer patients, developing the DeepSurv deep learning framework, which achieved a C-index of 0.824. YUAN et al. [15] demonstrated the superiority of an SVM classifier, developed using the ResNet-3D algorithm, over conventional contrast-enhanced CT scans in predicting the risk of synchronous peritoneal metastasis in colorectal cancer, exhibiting enhanced precision (94.1%) and AUC (0.922).

However, as the complexity of deep convolutional neural networks (CNNs) increases, there is a potential decline in training accuracy. To address this challenge, the incorporation of deep neural networks with residual connections has been proposed to bolster the learning efficiency of the network. This study employs the ResNet-34 model, augmented with residual connections, as the cornerstone for extracting features through deep transfer learning. The objective is to assess the efficacy of integrating MRI-based deep transfer learning, radiomics features, and clinical data in the accurate preoperative delineation between stages T2 and T3 of rectal cancer.

Materials and methods

Study subjects

This study is a retrospective analysis, which encompasses the complete medical records of 983 patients diagnosed with rectal cancer at the First Affiliated Hospital of Wannan Medical College between January 2018 and December 2022. The inclusion criteria were as follows: (1) All patients had undergone radical resection for rectal cancer, with postoperative pathology confirming the disease as either stage T2 or T3; (2) None of the patients had a history of any pelvic surgery, nor had they received pelvic radiotherapy, chemotherapy, or targeted therapy prior to the study; (3) Comprehensive preoperative MRI scans, clinical, and pathological data were available for all patients; (4) The MRI scans were of high quality, deemed suitable for detailed observation and analysis. Patients with benign or malignant tumors in the pelvic region or elsewhere were excluded from this study. According to these criteria, a total of 361 patients were selected for inclusion in the study, comprising 100 patients with stage T2 and 261 with stage T3 rectal cancer, including 236 males and 125 females. The age range of the participants was 22 to 88 years, with a mean age of 63.3 years (standard deviation = 10.7 years). These patients were then randomly assigned in a 7:3 ratio to either the training set (252 patients) or the test set (109 patients). The process of patient selection is depicted in the patient selection flow chart (as illustrated in Fig. 1).

Fig. 1.

Patient Selection Flowchart. A total of 983 patients diagnosed with rectal cancer were included. Exclusion criteria: patients with a history of pelvic surgery, or those who have received pelvic radiotherapy, chemotherapy, or targeted therapy.Finally, 361 eligible patients were included in the study, divided into a training set (252 patients) and a test set (109 patients)

This investigation was conducted in strict accordance with the ethical guidelines set forth in the Declaration of Helsinki. The Ethics Committee of Wannan Medical College approved the study protocol (Ethics Approval Number: (2023) Ethical Review No. (125)). As this study is a retrospective analysis, the requirement for informed consent was waived by the Ethics Committee.

Collection of clinical data

This investigation meticulously compiled essential demographic and clinical data for participants, encompassing gender, age, and post-surgical T staging. The delineation of postoperative T stages was conducted in strict adherence to the guidelines outlined in the 8th edition of the Cancer Staging Manual by the American Joint Committee on Cancer (AJCC) [16]. Specifically, T2 staging denotes the tumor’s invasion into but not beyond the muscularis propria, while T3 staging indicates the tumor’s extension beyond the muscularis externa into the surrounding mesorectal adipose tissue. In the week leading up to surgery, peripheral venous blood samples were collected from the subjects for precise quantitative assessments of Carcinoembryonic Antigen (CEA) and Carbohydrate Antigen 19 − 9 (CA19-9), utilizing the VIDAS fully automated immunoassay system. The established cutoff values for these analyses were defined as 5ng/ml for CEA and 37 U/ml for CA19-9, respectively.

MRI scanning protocol

Preoperative rectal MRI evaluations were conducted on all participants using a GE Signa HDxt 3.0 Tesla MRI scanner equipped with an 8-channel phased-array coil. Protocol for patient preparation required fasting and abstention from fluids for six hours preceding the scan, completion of urination and bowel prep, and the application of adequate abdominal compression to mitigate respiratory motion artifacts. Detailed scanning parameters were as follows: For axial T2-weighted imaging (T2WI), parameters were set to a repetition time (TR)/echo time (TE) of 4000 ms/80 ms, a field of view (FOV) of 240 mm × 240 mm, slice thickness of 3 mm, an interslice gap of 1 mm, a matrix size of 384 × 320, and a number of excitations (NEX) of 4. In diffusion-weighted imaging (DWI), the TR/TE was 5000 ms/88 ms, with a slice thickness of 4.0 mm, an FOV of 216 mm × 288 mm, a matrix size of 128 × 130, diffusion b values set at 0 and 1000 s/mm², and an NEX of 10. Measurements were conducted on the high-resolution T2WI to determine the tumor’s longest diameter (LD), while the apparent diffusion coefficient (ADC) values were quantified using the ADC maps generated from DWI.

Image segmentation and radiomic feature extraction

Upon retrieval of image datasets in DICOM format from the imaging repository, this investigation employed the grayscale normalization function within ITK-SNAP software (version 3.6.0) for image standardization. In axial T2WI, The Regions of Interest (ROIs) were manually delineated along the tumor margins using ITK-SNAP software by two radiologists, each with over 10 years of clinical radiology experience and extensive practical expertise in tumor imaging. This ensured that the tumor boundaries in each slice were accurately marked, thereafter amalgamating these delineations to configure the volume of interest (VOI) (illustrated in Fig. 2). Following this, the Pyradiomics platform was utilized to extract a comprehensive suite of 1562 unique radiomic features per image, These features included shape features, texture features, and intensity features. To ensure the uniformity and dependability of the extracted data, a rigorous consistency assessment was conducted on the radiomic features identified by both practitioners, employing the Intraclass Correlation Coefficient (ICC). Only features demonstrating an ICC value above 0.80 were preserved, thereby safeguarding the precision and robustness of the subsequent analysis.

Fig. 2.

Image Segmentation. A depicts the two-dimensional ROIs based on lesion delineation, B shows the three-dimensional VOI generated from the fusion of ROIs, and C presents the distribution of radiomic features

Deep transfer learning for feature extraction

Transfer learning draws inspiration from the human capability to apply acquired knowledge to novel contexts. Within the domain of deep learning, this typically involves the application of models, pre-trained on extensive datasets, to tasks with relatively limited data. The principal benefit of transfer learning lies in its ability to utilize the sophisticated feature representations gleaned from the pre-trained models, thus expediting the learning process and enhancing the efficacy of new tasks, particularly under conditions of sparse labeled data. By transferring pre-existing knowledge, this method adeptly addresses the challenges posed by learning in data-deficient environments.

This research opts for ResNet34 as the cornerstone model for deep transfer learning in feature extraction. Comprising 34 convolutional layers, ResNet34 achieves an optimal balance between high performance and computational efficiency. Its structure is segmented into four primary blocks of layers, each containing a series of residual blocks: three in Layer 1, four in Layer 2, six in Layer 3, and three in Layer 4. Within each residual block are several convolutional layers, employing batch normalization and ReLU activation functions. These layers are interconnected by skip connections, which directly add the input to the output, as depicted in Fig. 3. The pre-training of this architecture on the voluminous ImageNet dataset, which encompasses millions of images across 1,000 categories, enables the ResNet34 model to acquire a rich tapestry of image feature representations. For the purposes of this study, the network’s architecture beyond the Fully Connected layer is excised. Inputting all data into the pre-trained ResNet34 model allows for the extraction of 512 deep transfer learning features for each case from the subsequent Average Pooling layer, enhancing the precision and efficiency of feature representation in new domains.

Fig. 3.

Schematic of Deep Transfer Learning Based on ResNet34. Left: The overall structure of the ResNet-34 model, which includes 34 convolutional layers and residual connections.Right: The detailed structure of each residual block, showing the convolutional layers, batch normalization layers, and ReLU activation functions within each block.Bottom: The model’s input and output features, with the input being the raw imaging data and the output being 512 deep learning features

Construction of the model

First, we performed Spearman rank correlation analysis on the extracted features to evaluate the correlation between each feature and the target variable (rectal cancer staging). Features with high correlation to the target variable were selected, while those with low correlation were excluded. Based on this preliminary screening, we further applied the Least Absolute Shrinkage and Selection Operator (LASSO) regression techniques for feature selection. LASSO regression, by introducing L1 regularization in the regression model, can automatically select and shrink the coefficients of unimportant features, thus achieving feature reduction and selection.Following this, distinct predictive models were developed using a suite of diverse machine learning algorithms: Logistic Regression (LR), Random Forest (RF), Decision Trees (DT), and Support Vector Machines (SVM). A comparative analysis was conducted to identify the algorithm that delivered superior performance. The most effective model was subsequently amalgamated with clinical data to formulate a holistic predictive model. To provide a visual representation of this model’s predictive capabilities, a Nomogram and calibration curves were created as evaluative tools to gauge the accuracy of the model’s predictions. The workflow for model construction is depicted in Fig. 4.

Fig. 4.

The Overall Workflow Diagram of This Study. ROIs Segmentation: Manual delineation of 2D ROIs to generate 3D VOI.Feature Extraction: Extraction of radiomic and deep learning features from the 3D VOI.Feature Selection: Feature selection using Spearman rank correlation analysis and LASSO regression.Model Construction and Evaluation: Construction of predictive models using various machine learning algorithms and performance evaluation

Statistical analysis

In this investigation, the comparison of categorical variables was performed using either Chi-square tests or Fisher’s exact tests to evaluate statistical significance. Independent risk factors were elucidated through both univariate and multivariate logistic regression analyses, with the calculation of their odds ratios and corresponding 95% confidence intervals. To identify the radiomic features with the highest predictive power, we employed dimensionality reduction techniques, specifically Spearman’s rank correlation and LASSO regression techniques. Moreover, predictive models were developed leveraging a suite of diverse machine learning algorithms, including LR, RF, DT, and SVM. The diagnostic performance of these models was assessed by determining the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC), along with sensitivity and specificity metrics. The threshold for statistical significance across all analyses was established at P < 0.05.

Results

Comparison of clinical data between training set and test set

In the training set of this study, a total of 252 patients were included, with 69 patients in stage T2 and 183 patients in stage T3. The gender distribution was 168 males and 84 females. The age range was between 32 and 88 years, with an average age of 63.2 years (standard deviation: 10.3 years). In the test set, there were 109 patients, including 31 patients in stage T2 and 78 patients in stage T3, with 68 males and 41 females. The age range was from 22 to 85 years, with an average age of 63.6 years (standard deviation: 11.6 years). When comparing the clinical data between the training set and the test set, no statistically significant differences were observed (all P > 0.05). Detailed data can be found in Table 1.

Table 1.

Comparison of Clinical Data between Training Set and Test Set

Clinical Data	Training Set(n = 252)	Test Set(n = 109)	Statistical Values	P Values
Gender (example)			0.616	0.432
Male	168(66.7)	68(62.4)
Female	84(33.3)	41(37.6)
Age (years old)			0.134	0.715
≧ 60	160(63.5)	67(61.5)
<60	92(36.5)	42(38.5)
LD(cm)			4.741	0.093
<3	29(11.5)	22(20.2)
3 ~ 5	176(69.8)	68(62.4)
≧ 5	47(18.7)	19(17.4)
CEA(ng/mL)			0.579	0.447
>5	107(42.5)	51(46.8)
≤ 5	145(57.5)	58(53.2)
CA19-9(U/mL)			1.968	0.161
>37	26(10.3)	17(15.6)
≤ 37	226(89.7)	92(84.4)

Construction of machine learning models

In this research, we harvested a comprehensive set of 1562 radiomic features and 512 deep transfer learning features from tumor data. we meticulously selected 24 features of utmost predictive significance, comprising 4 radiomic and 20 deep transfer learning features. It is important to highlight that the deep transfer learning features exhibited substantially greater significance in comparison to the radiomic features, as illustrated in Fig. 5. Leveraging these meticulously chosen 24 features, we developed predictive models employing four distinguished machine learning algorithms: LR, RF, DT, and SVM. Detailed assessments of each model’s performance are systematically presented in Tables 2and Fig. 6.

Fig. 5.

Showcases the 10-fold cross-validation procedure conducted via LASSO regression. In Fig. A, the 10-fold cross-validation technique was utilized to pinpoint the optimal hyperparameter λ (lambda) for the LASSO model. Here, the x-coordinate at which the model error is minimized aligns with the optimal λ value, as denoted by the dashed line. Figure B illustrates the trajectories of feature coefficients as a function of λ, with each colored line indicating the evolution of a coefficient for a specific feature. Utilizing the λ value identified in Fig. A (marked by the dashed line), features possessing non-zero coefficients were selected for inclusion in the ultimate model build. Figure 5 C reveals these carefully chosen optimal features along with their respective weights. Conclusively, Fig. 5D provides a heatmap depicting the correlation matrix between the selected optimal features and T stages, elucidating the degrees of correlation both within the features themselves and between the features and T stages, thereby offering a nuanced view of their interrelations

Table 2.

Diagnostic efficacy of four machine learning models

	AUC(95%CI)	Sensitivity(%)		Specificity(%)
Training Set
LR	0.867(0.821 ~ 0.914)		83.1	72.5
RF	0.834(0.781 ~ 0.888)		68.9	87.0
DT	0.900(0.858 ~ 0.943)		71.0	92.8
SVM	0.944(0.911 ~ 0.977)		89.6	88.4
Test Set
LR	0.847(0.744 ~ 0.950)		84.9	78.3
RF	0.803(0.705–0.901)		76.7	78.3
DT	0.842(0.752–0.933)		74.4	87.0
SVM	0.910(0.816-1.000)		100	82.6

Fig. 6.

Illustrates the efficacy of models developed through four distinct machine learning algorithms, as evaluated by their ROC curve metrics. Figure A showcases the ROC curve outcomes for these models within the training set, whereas Fig.B delineates their performance metrics on the test set

Model comparison

The SVM algorithm was identified as the most optimal model. Using the SVM algorithm, we developed predictive models based on both deep transfer learning features and traditional radiomics features. The results demonstrated that the deep transfer learning model outperformed the traditional radiomics model in both the training and test set (P < 0.05), as shown in Table 3.

Table 3.

Comparison of models based on deep transfer learning features and traditional Radiomics features

Group	Model	AUC(95%CI)	Sensitivity (%)	Specificity (%)	Z	P
Training set	Deep Transfer Learning Model	0.938(0.900 ~ 0.975)	95.1	72.5	6.848	<0.001
	Traditional Radiomics Model	0.667(0.592 ~ 0.743)	61.7	68.1
Test set	Deep Transfer Learning Model	0.885(0.763 ~ 1.000)	98.8	69.6	2.110	0.035
	Traditional Radiomics Model	0.708(0.586 ~ 0.831)	83.7	56.5

Analysis of clinical risk factors

In this investigation, initial univariate logistic regression analyses were performed on a series of clinical parameters, namely gender, age, maximum tumor LD, CEA, and CA19-9. Following this, variables exhibiting statistical significance (P < 0.05) in the univariate phase were advanced into a multivariate logistic regression analysis aimed at pinpointing independent risk factors crucial for distinguishing between T2 and T3 stages of rectal cancer. The univariate analysis outcomes indicated statistically significant disparities in LD, CEA, and CA19-9 levels (P < 0.05), with these parameters presenting elevated values in T3 stage rectal cancer relative to T2 stage. Further multivariate analysis discerned LD and CEA as independent predictors for differentiating between the T2 and T3 stages of rectal cancer. For a comprehensive presentation of the findings, refer to Table 4.

Table 4.

Analysis of clinical risk factors for Predicting T2 versus T3 staging in rectal Cancer

Clinical Factors	Univariate logistic regression	P Value	Multivariate logistic regression	P Value
	OR(95%CI)		OR(95%CI)
Gender	1.093(0.610 ~ 1.960)	0.764	—	—
Age	0.772(0.428 ~ 1.391)	0.389	—	—
LD		0.003		0.020
LD(1)	2.352(1.058 ~ 5.227)	0.036	2.029(0.885 ~ 4.648)	0.094
LD(2)	7.840(2.441 ~ 25.491)	0.001	5.684(1.686 ~ 19.161)	0.005
CEA	3.640(1.917 ~ 6.911)	<0.001	2.778(1.431 ~ 5.394)	0.003
CA19-9	10.759(1.429 ~ 81.019)	0.021	7.050(0.910 ~ 54.629)	0.062

Comprehensive predictive model

In this study, we selected the SVM as our definitive predictive model due to its superior performance on both training and test datasets. Integrating clinical independent risk factors, specifically LD and CEA, we developed an encompassing predictive model. Additionally, a nomogram alongside the ROC and calibration curves were constructed to evaluate the model’s predictive accuracy, as shown in Fig. 7. This comprehensive model displayed excellent predictive efficacy in the training set, with an AUC of 0.946 (95% CI: 0.914–0.977), sensitivity of 93.4%, and specificity of 85.5%. Similarly, in the test set, the model maintained outstanding performance, achieving an AUC of 0.920 (95% CI: 0.829–1.000), sensitivity of 100%, and specificity of 82.6%.

Fig. 7.

Presents multiple key visuals of the comprehensive predictive model in this study. Figure A displays the nomogram of the model, which serves as a visual tool for predicting outcomes. Figure B shows the ROC curves of the model on the training and validation sets, evaluating its diagnostic efficacy. Figures C and D depict the calibration curves of the model on the training and test sets, respectively, illustrating the accuracy of the model’s predictions

Discussion

Our study capitalized on MRI-based deep transfer learning techniques integrated with radiomic features, utilizing LR, RF, DT, and SVM machine learning algorithms to construct four predictive models effectively. Among these, the SVM model exhibited superior performance. By deploying the SVM model as the conclusive output and incorporating clinical independent risk factors, we formulated an encompassing predictive model. This model attained AUC values of 0.946 and 0.920 within the training and test datasets, respectively. Despite histopathology remaining the gold standard for tumor staging, the acquisition of tumor tissue samples predominantly depends on invasive surgical procedures or biopsies, potentially leading to inherent sampling errors and observer variation in pathological analysis. Research highlights that between 7 and 15% of patients encounter uncertainty in pathological diagnosis, emphasizing the critical need for imaging examinations to more comprehensively evaluate tumor heterogeneity [17, 18]. In this context, non-invasive MRI imaging technology assumes a pivotal role in the diagnostic staging of rectal cancer [19–22]. Nevertheless, the present constraints of imaging modalities open avenues for advanced image feature extraction and analysis methodologies. Employing high-throughput analysis at the voxel level of MRI images, radiomics, and deep learning facilitate the non-invasive prediction of rectal cancer’s pathological staging, extracting profound features including morphology, histology, and functionality. This methodology, being more objective than conventional approaches, has demonstrated gratifying predictive results. Our 3D CNNs predictive model, predicated on axial high-resolution T2WI, exhibited significant diagnostic effectiveness across both training and test datasets, affirming the model’s robustness as an efficacious, non-invasive, pre-surgical technique for forecasting rectal cancer staging.

ResNet is currently one of the most acclaimed architectures in the deep learning landscape. Research employing the ResNet-50 model has effectively distinguished between glioblastomas and solitary brain metastases. This achievement was realized through the development of a model utilizing 2D CNNs, which attained an AUC of 0.835 in external test sets [23]. Furthermore, some researchers have incorporated the ResNet-34 architecture into the convolutional blocks of the U-Net network to classify the tumor’s enhanced, non-enhanced, and necrotic regions. This novel method has markedly enhanced model efficiency while also expediting the speed of model convergence and deepening the training process [24]. These studies have leveraged the ResNet architecture to develop 2D classification models. Contrarily, our investigation employed a distinct approach by executing random cropping around tumor contours instead of inputting full images. This technique efficiently refined the model’s depth and performance by extracting pertinent tumor features and eliminating superfluous ones. Prior studies involved processing MRI images that were converted into JPG format scene images. Although this method showed effectiveness, it neglected the inter-layer connectivity inherent in MRI images, thereby constraining the scope for model enhancement. As a result, we transitioned from 2D CNNs to fully embracing 3D CNNs, aiming to thoroughly leverage the correlations present in MRI images across three-dimensional spaces, encompassing vertical, horizontal, and axial dimensions. This innovative tactic is anticipated to furnish the medical image processing domain with more detailed and accurate information, thereby offering more robust support for disease diagnosis and analytical processes. Furthermore, by filtering out irrelevant voxels, the 3D network not only bolsters the precision of model classification but also simplifies the training process [25, 26].

Tumor marker detection is characterized by its minimally invasive approach and the potential for frequent short-term assessments, making strategic utilization of tumor marker testing invaluable for the screening, diagnosis, staging, and prognosis of cancer [27–30]. Research has identified over ten serum tumor markers associated with colorectal cancer, among which CEA and CA19-9 are the most widely utilized. CEA, a high-molecular-weight glycoprotein produced by normal rectal cells, functions as an intercellular adhesion molecule facilitating the aggregation of colorectal cancer cells. Conversely, CA19-9, a high-molecular-weight glycolipid, primarily influences cell adhesion functions, playing a pivotal role in the progression of tumors [31]. This study observed elevated levels of CEA and CA19-9 in patients with stage T3 colorectal cancer compared to those in stage T2, suggesting that higher levels of these markers may indicate enhanced proliferative capabilities of tumor cells, lower levels of differentiation, and greater malignancy severity. The significantly higher malignancy in stage T3 colorectal cancer as opposed to T2 implies a broader disparity in differentiation levels. Previous research corroborates the utility of CEA and CA19-9 in diagnosing colorectal cancer, prognosticating outcomes, and monitoring recurrence [27, 28]. LIN et al. [32]. , in their investigation using radiomics nomograms for the preoperative prediction of colorectal cancer T staging, demonstrated that both CEA (OR = 4.08, 95%CI:1.859.00) and CA19-9 (OR = 5.83, 95%CI:1.3325.62) were statistically significant in univariate analyses (P < 0.05), with CEA emerging as an independent risk factor in multivariate analysis (P = 0.044), findings that mirror our own. Earlier studies have noted a correlation between tumor invasiveness and size [33], suggesting that larger tumors possess greater invasive potential, deeper infiltration into the intestinal wall, and subsequently, a higher T staging. In our analysis, the LD of tumors in the T2 stage was found to be smaller than in T3, with LD proving to be an independent risk factor for distinguishing between these stages (OR = 5.117, 95%CI:1.159 ~ 22.584), a consistency with prior reports.

In this study, we utilized the ResNet-34 model with residual connections for feature extraction to achieve deep transfer learning. This approach has significant advantages in complex feature extraction, especially when handling high-dimensional imaging data. Compared to traditional methods, the ResNet-34 model, with its deep network structure and residual connections, is more effective at capturing subtle features in imaging data, thereby improving the model’s predictive accuracy.Traditional radiomics feature models rely on predefined imaging features, which are often manually selected by experts and may suffer from feature redundancy and information loss. Although these features can provide some predictive ability in certain cases, their performance is often limited by the methods of feature selection and extraction. In our study, the traditional radiomics feature model showed AUC of 0.667 and 0.708 in the training and validation groups, respectively, indicating lower efficacy.Predictive models based on clinical features (such as age, gender, tumor markers, etc.), while important in disease diagnosis and prognosis, have lower predictive accuracy when used alone. In our study, the clinical model demonstrated AUC of 0.694 and 0.653 in the training and validation set, respectively, indicating lower predictive performance.

While this study yields significant insights, it is not without its limitations. Firstly, due to the data being derived from a single center, the population distribution and study scope are somewhat narrow. Consequently, our findings may be influenced by specific regional and demographic traits, limiting their applicability across diverse regions or populations. To augment the generalizability and reliability of our findings, future research should be conducted in a multicenter setting to validate and broaden our observations. Secondly, this study utilized a retrospective analysis methodology, potentially introducing selection bias. Despite endeavors to mitigate bias within the study design, the inherent constraints of retrospective studies may still impinge on the interpretation of our findings.Furthermore, our advancements in employing tumor markers such as CEA and CA19-9, alongside MRI imaging characteristics for colorectal cancer staging, introduce a novel paradigm for non-invasive diagnosis. Nevertheless, forthcoming research should investigate an expanded spectrum of biomarkers and the potential amalgamation of cutting-edge imaging technologies to elevate the precision and efficiency of colorectal cancer staging. Moreover, the incorporation of novel machine learning models and algorithms, particularly for analyzing complex medical imaging data, could yield more profound insights into early tumor diagnosis and treatment strategies. In conclusion, future research endeavors should focus on enlarging sample sizes, adopting prospective study designs, and delving into a broader range of biomarkers and imaging parameters. This approach will provide a more thorough and accurate foundation for diagnosing and treating colorectal cancer.

Conclusions

Synthesizing the results of this investigation, we adeptly merged MRI-based deep transfer learning techniques with radiomic features, utilizing four distinct machine learning algorithms—LR, RF, DT, and SVM—to create four predictive models. The outcomes revealed that each algorithm displayed significant predictive efficacy, with the SVM algorithm notably excelling. By integrating clinical risk factors into the predictive framework, we’ve crafted a comprehensive model that can non-invasively and precisely forecast the T2 and T3 staging of colorectal cancer before surgery. This innovation serves as a tailored decision-support tool for the preoperative evaluation of colorectal cancer patients, offering essential insights for clinicians to formulate more individualized clinical treatment strategies.

Abbreviations

LASSO

Least Absolute Shrinkage and Selection Operator

LR

Logistic Regression

RF

Random Forest

DT

Decision Tree

SVM

Support Vector Machine

AUC

The Area Under The Curve

NCCN

National Comprehensive Cancer Network

TME

Total Mesorectal Excision

nCRT

Neoadjuvant Chemoradiotherapy

CNNs

Convolutional Neural Networks

AJCC

American Joint Committee on Cancer

CEA

Carcinoembryonic Antigen

CA19-9

Carbohydrate Antigen 19-9

T2WI

T2-Weighted Imaging

RT

Repetition Time

ET

Echo Time

FOV

Field Of View

NEX

Number Of Excitations

LD

Longest Diameter

ADC

Apparent Diffusion Coefficient

ROIs

Regions Of Interest

VOI

Volume Of Interest

ICC

Intraclass Correlation Coefficient

ET

Echo Time

Author contributions

LF, HW, YW, SW, JZ, XZ contributed to the study conception and design. The frst draft of the manuscript was written by LF and all authors commented on previous versions of the manuscript. LF, HW, YW, SW, JZ, XZ read and approved the final manuscript.

Funding

This work was supported by The open project of Anhui Province Key Laboratory of Cancer Translational Medicine, Bengbu Medical University(NO.KFKT202307); Anhui provincial Department of Education university research project (Grant No. 2023AH051763).Wannan Medical College Key Research Project Fund at the School Level (No.WK2023ZZD09);Wannan Medical College Middle-aged and Young Project Research Fund (No. WK2023ZQNZ53).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This investigation was conducted in strict accordance with the ethical guidelines set forth in the Declaration of Helsinki. The Ethics Committee of Wannan Medical College approved the study protocol (Ethics Approval Number: (2023) Ethical Review No. (125)). As this study is a retrospective analysis, the requirement for informed consent was waived by the Ethics Committee.

Consent for publication

Not Applicable.

Competing interests

The authors declare no competing interests.