BPG is committed to discovery and dissemination of knowledge
Retrospective Study Open Access
Copyright: ©Author(s) 2026. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. No commercial re-use. See permissions. Published by Baishideng Publishing Group Inc.
World J Gastrointest Oncol. Nov 15, 2026; 18(11): 123187
Published online Nov 15, 2026. doi: 10.4251/wjgo.123187
Establishment and validation of a multimodal model using deep learning for the recognition of early colorectal cancer
Wei-Hua Dong, Xin Ling, Xiao-Dong Chen, Jia-Ping Qian, Ying Jin, Yi-Jie Gu, Department of Gastroenterology, Suzhou Ninth People’s Hospital, Suzhou 215000, Jiangsu Province, China
ORCID number: Wei-Hua Dong (0009-0000-6248-9874); Ying Jin (0009-0001-2933-2653); Yi-Jie Gu (0000-0001-7129-0007).
Co-first authors: Wei-Hua Dong and Xin Ling.
Co-corresponding authors: Ying Jin and Yi-Jie Gu.
Author contributions: Dong WH designed the study and drafted the manuscript; Ling X and Chen XD collected the clinical data and image materials used in the study and performed the statistical analysis; Qian JP, Gu YJ and Jin Y designed the study and created all the figures in the manuscript. All authors read and approved the manuscript. Dong WH and Ling X contributed equally to this work as co-first authors. This study involves a multidisciplinary workflow integrating clinical data collection, endoscopic image processing, and deep learning model development. Gu YJ and Jin Y contribute distinct and complementary expertise that is essential to the successful execution of this project. Gu YJ serves as the lead investigator for data processing and algorithmic development. He is responsible for the construction and optimization of the deep learning models, including data preprocessing, feature extraction, model training, and performance evaluation. He also developed the multimodal nomogram and performed all statistical analyses reported in this manuscript. His technical expertise ensures the reproducibility and methodological rigor of the computational aspects of this study. Jin Y, as the senior gastroenterologist, is primarily responsible for the clinical conception, study design, and interpretation of endoscopic findings. She oversees the clinical data acquisition, ensures the accuracy of pathological correlations, and coordinates the collaborative efforts among all team members, including endoscopists, pathologists, and data scientists. Her clinical leadership guarantees that the model outputs are clinically relevant and interpretable. Both investigators bear equal responsibility for the overall scientific integrity of the study.
AI contribution statement: Portions of this manuscript were edited using AI tools solely for language refinement. The authors carefully reviewed and verified all AI-assisted outputs and take full responsibility for the scientific content of the manuscript.
Supported by Youth Scientific Research Foundation of the Ninth People’s Hospital of Suzhou, No. YK202405 and No. YK202506.
Institutional review board statement: The study protocol was approved by the Ethics Committee of the Ninth People's Hospital of Suzhou (Approval No. YJKY2025-012) in accordance with the ethical principles of the Declaration of Helsinki of the World Medical Association and the International Ethical Guidelines for Biomedical Research Involving Human Subjects of the Council for International Organizations of Medical Sciences.
Informed consent statement: All patients provided written informed consent.
Conflict-of-interest statement: All the authors report no relevant conflicts of interest for this article.
Data sharing statement: The datasets generated and/or analyzed during the current study are not publicly available due to institutional policies and patient privacy protection, but are available from the corresponding author upon reasonable request.
Corresponding author: Yi-Jie Gu, Department of Gastroenterology, Suzhou Ninth People’s Hospital, No. 2666 Lvdang Road, Wujiang District, Suzhou 215000, Jiangsu Province, China. guyijie1997@163.com
Received: May 11, 2026
Revised: July 21, 2026
Accepted: September 23, 2026
Published online: November 15, 2026
Processing time: 178 Days and 12.9 Hours

Abstract
BACKGROUND

Colorectal cancer is a leading cause of cancer-related mortality, yet distinguishing early colorectal cancer (ECC) from non-cancerous lesions under white-light endoscopy remains challenging. We therefore developed a multimodal model integrating clinical variables with deep learning features from white-light images to differentiate ECC from non-cancerous lesions.

AIM

To develop and validate a multimodal model leveraging convolutional neural networks (CNNs) to detect ECC.

METHODS

Data from patients with colorectal lesions treated at Suzhou Ninth People’s Hospital between January 2020, and March 2025, were analyzed retrospectively, including clinical data, endoscopic images, and pathological results. Clinically significant variables were identified through univariate and multivariate analyses, while endoscopic image features were optimized using CNNs. A multimodal model was established by integrating these parameters. Model performance was evaluated using an area under the curve (AUC), calibration curves, and decision curve analysis (DCA).

RESULTS

A total of 576 cases, including 180 ECC and 396 non-cancerous lesions with 1152 images, were divided into training and validation sets (7:3). In the training set, multivariate analysis identified lesion length [odds ratio (OR) = 11.377, 95% confidence interval (CI): 6.839-18.928, P < 0.001] and rectal location (OR = 2.407, 95%CI: 1.190-4.868, P = 0.015) as independent risk factors for ECC. The ensemble deep learning prediction derived from VGG19, ResNet50, and InceptionV3 models yielded a sensitivity of 0.85, specificity of 0.93, and AUC of 0.89. The multimodal model, incorporating lesion location, lesion length, and ensemble deep learning prediction, achieved a sensitivity of 0.80, specificity of 0.94, and an AUC of 0.95 (95%CI: 0.91-0.99) in the validation set. Calibration curves confirmed favorable agreement, and DCA demonstrated net clinical benefit.

CONCLUSION

This multimodal model effectively distinguishes non-cancerous lesions from ECC, demonstrating strong diagnostic performance and potential to improve early cancer detection by endoscopists.

Key Words: Early colorectal cancer; Deep learning; Ensemble learning; Multimodal model; White-light endoscopy; Computer-aided diagnosis

Core Tip: We developed and validated a multimodal model combining clinical variables (lesion length and location) with an ensemble deep learning prediction from three convolutional neural networks (VGG19, ResNet50, and InceptionV3) to distinguish non-cancerous lesions from early colorectal cancer using only white-light endoscopic images. The model achieved a sensitivity of 0.80, specificity of 0.94, and area under the curve of 0.95 in the validation set, outperforming both junior and senior endoscopists in specificity and accuracy. This practical, hardware-independent approach offers a promising tool to assist endoscopists in improving early cancer detection during routine colonoscopy.



INTRODUCTION

Colorectal cancer is the third most common malignant tumor worldwide and is also one of the most prevalent malignancies in China[1-3]. Colonoscopy serves as a common diagnostic method for early colorectal cancer (ECC) screening, but assessing the nature and depth of lesion remains challenging. In the process of “adenoma-cancer”, methods for evaluating lesion characteristics include the NICE classification, JNET classification, pit pattern classification, and the WASP system[4-6]. The implementation of these classifications presents dual challenges: A significant learning burden for endoscopic trainees and a dependency on advanced hardware, including magnifying or image enhanced endoscopy. Therefore, a more accessible and simplified endoscopic diagnostic model is urgently needed.

ECC has garnered increasing attention in recent literature. ECC is defined as carcinoma confined to the mucosal and submucosal layers, regardless of size or lymph node involvement. When the depth of invasion is less than 1000 µm, with negative margins and without vascular or lymph node involvement, ECC can be removed by endoscopic curative resection without surgery[7]. However, ECC is usually difficult to diagnose under the white light model. Previous studies indicate that endoscopists' diagnostic accuracy for characterizing colorectal polyps under white-light and narrow-band imaging ranges from 45% to 80%, with considerable variation among practitioners[5,8,9]. Artificial intelligence technology, represented by convolutional neural networks (CNN), demonstrate promising applications in image recognition[10,11].

Several reports have demonstrated the application of deep learning in identifying colorectal polyps, adenomas, and colorectal cancers[12-14]. However, a significant research gap persists in the development of diagnostic frameworks relying exclusively on white light endoscopic images and basic clinical variables to discriminate between non-cancerous lesions and early-stage colorectal cancer. To address this limitation, this study constructed and validated a multimodal model integrating clinical data and white-light endoscopic images for the preoperative differentiation of non-cancerous lesions and ECC.

MATERIALS AND METHODS
Research subject

This study retrospectively collected data from Suzhou Ninth People’s Hospital between January 1, 2020, and March 1, 2025, all lesions underwent endoscopic treatment and were diagnosed as non-cancerous lesions or ECC by pathologists. The images were taken from the 290 system of Olympus Corporation in Japan. Image exclusion criteria: Images with blurring, defocusing, bleeding, fecal coverage, poor lighting, or reflection artifacts were excluded from analysis. Non-cancerous lesions included inflammatory lesions, serrated lesions, and adenomas. ECC referred to tumors tissue confined to the mucosal layer and submucosa. This study was approved by the hospital ethics committee (Approval No. YJKY2025-012). Basic patient information (gender, age); lesion characteristics (location, length); white-light endoscopic images; and HE staining results were collected.

Dataset classification

Stratified random sampling was performed using the R package caret with the random seed set to 1234. The dataset was partitioned into training and validation sets at a 7:3 ratio while preserving the original distribution of non-cancerous lesions vs ECC. The training set was utilized for feature extraction and model development, whereas the validation set was reserved for evaluating the diagnostic performance of the established model.

Research methods

Establishment of a logistic regression model based on clinical data: In the training set, based on the basic information of patients and the characteristics of lesions, variables with statistical differences were screened through univariate analysis and multivariate analysis, and a logistic regression model was established. Its sensitivity, specificity, accuracy rate and area under the receiver operating characteristic curve (AUROC) were calculated.

Establishment of a white-light endoscopic prediction model based on deep learning: Preprocessed white-light endoscopic images from the training set were processed with PyTorch (version 2.3.1). The preprocessing pipeline included: Resizing to 256 pixels × 256 pixels, random cropping (retaining 60%-100% of original area with aspect ratio 0.8-1.0) followed by resizing to 224 pixels × 224 pixels, random horizontal flipping (50% probability), random rotation within ± 20 degrees, random brightness adjustment (factor 0.5-1.5), and random contrast adjustment (factor 0.5-1.5). Images were converted from PIL or numpy format to tensors and normalized to[0,1], followed by standardization to[-1,1] to enhance model generalization and prevent overfitting.

Sensitivity was defined as the proportion of pathologically confirmed ECC cases that were correctly identified by the model [true positives/(true positives + false negatives)]. Specificity was defined as the proportion of pathologically confirmed non-cancerous cases that were correctly identified by the model [true negatives/(true negatives + false positives)]. Accuracy was calculated as (true positives + true negatives)/total cases.

Three CNN architectures, including VGG19, ResNet50, and InceptionV3, were fine-tuned using ImageNet pre-trained weights for binary classification of non-cancerous lesions vs ECC.

Model optimization employed the Adam optimizer with adaptive learning rate adjustment for parameters in the fully-connected layers. All three CNN models were trained using the Adam optimizer with a fixed learning rate of 0.001, a batch size of 16, and a maximum of 50 training epochs. Early stopping was applied based on validation loss to prevent overfitting. The same hyperparameter settings were used across all models to ensure a fair comparison.

Performance metrics including sensitivity, specificity, and accuracy were evaluated for each individual network. To achieve the required high sensitivity for clinical applications, an ensemble decision rule was established: Lesions classified as ECC by at least one of the three models were designated as malignant, while all remaining cases were categorized as non-cancerous lesions. The output of this ensemble rule was defined as the “ensemble deep learning prediction”. In this study, two images from different angles were selected for each lesion. In the final multimodal model, the predicted probabilities of the two images were averaged to obtain the final prediction probability for each case.

Establishment and validation of the multimodal model: These clinically significant parameters extracted from the above models were incorporated into a nomogram model to construct the multimodal model. Diagnostic performance, including sensitivity, specificity, and accuracy, was evaluated in the validation cohort. Receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and calibration curves were generated to assess model discrimination, clinical utility, and prediction accuracy, respectively. The validation set was independently evaluated using six distinct approaches: Predictions from individual VGG19, ResNet50, and InceptionV3 models; logistic regression analysis; the ensemble deep learning model; and the integrated multimodal model. Concurrently, endoscopic assessments were performed by two clinical groups: Junior endoscopists (less than three years of experience) and senior endoscopists (over three years of experience). In the validation set, the multimodal model and the endoscopists evaluated the lesions under identical conditions. The information provided included patient sex, age, lesion size, lesion location, and two static endoscopic images. Sensitivity, specificity, accuracy, and area under the curve (AUC) values were systematically compared across all evaluation methods. The complete analytical workflow is illustrated in Figure 1.

Figure 1
Figure 1 Flowchart of this study. A: Dataset composition and stratified random split into training and validation sets (7:3 ratio); B: Univariate and multivariate logistic regression analyses of clinical data in the training set to identify independent risk factors for early colorectal cancer; C: Deep learning model training using VGG19, ResNet50, and InceptionV3 architectures on endoscopic images from the training set, followed by Gradient-weighted Class Activation Mapping visualization for model interpretability; D: Construction of the multimodal nomogram integrating clinical variables and ensemble deep learning prediction; E: Comprehensive evaluation of the multimodal model using receiver operating characteristic curves, calibration curves, decision curve analysis, and comparison with endoscopists of varying experience levels. ROC: Receiver operating characteristic; DCA: Decision curve analysis; Grad-CAM: Gradient-weighted Class Activation Mapping.
Statistical analysis

SPSS 26.0 software was used for statistical analysis of the data, and PyTorch was used for image analysis. Normality tests were conducted on all continuous variables. Continuous variables that conformed to the normal distribution were expressed as mean ± SD, non-normally distributed continuous variables were expressed as the median and interquartile range (M, Q1-Q3). The comparison of continuous variables was conducted using the independent sample t-test. Categorical variables were expressed by the number of parameter, n (%), and Fisher’s exact test or χ2 test was used for categorical variables. Variables were selected for inclusion in logistic regression through univariate and multivariate analyses. The R language was used to draw nomogram, ROC curves, DCA, and calibration curves. P value < 0.05 was set as statistically significant.

RESULTS
Baseline data

This study included 576 lesions from Suzhou Ninth People’s Hospital. Among them, 383 were male (66.5%) and 193 were female (33.5%). The mean age of patients was 60.78 ± 10.40 years, and the mean lesion length was 1.20 ± 0.79 cm. ECCs accounted for 180 cases (31.3%), while non-cancerous lesions comprised 396 cases (68.7%). 492 lesions (85.4%) were located in the colon, and 84 lesions (14.6%) were located in the rectum.

Through stratified random sampling based on the pathology of lesions, the data were divided into training and validation sets in a 7:3 ratio. The training set comprised 404 cases, including 275 non-cancerous lesions and 129 ECCs. The validation set comprised 172 cases, including 121 non-cancerous lesions and 51 ECCs.

Demographic and endoscopic characteristics of both cohorts are presented in Table 1. No statistically significant differences were observed between the training and validation sets regarding age (P = 0.169), gender distribution (P = 0.277), lesion length (P = 0.212), pathological diagnosis (P = 0.589), or lesion location (P = 0.983).

Table 1 Basic information of training set and validation set, n (%).
Item
Training set
Validation set
Statistical value
P value
Sample size404172--
Age (years, mean ± SD)61.17 ± 10.4759.87 ± 10.21t = 1.3760.169
Lesion length (cm, mean ± SD)1.23 ± 0.811.13 ± 0.73t = 1.2490.212
Genderχ2 = 1.1800.277
    Female141 (34.9)52 (30.2)
    Male263 (65.1)120 (69.8)
Pathological diagnosisχ2 = 0.2920.589
    Non-cancerous275 (68.1)121 (70.3)
    Early cancer129 (31.9)51 (29.7)
Lesion locationχ2 < 0.0010.983
Colon345 (85.4)147 (85.5)
Rectum59 (14.6)25 (14.5)
Variable screening from clinical data

In the training set, the mean age of patients with non-cancerous lesions was 60.82 ± 10.85 years. 101 cases (36.7%) were female, the mean lesion length was 0.88 ± 0.46 cm, and 248 lesions (90.2%) were located in the colon. The mean age of patients with ECC was 61.91 ± 9.60 years. 40 cases (31.0%) were female. The mean lesion length was 1.95 ± 0.92 cm, and 97 lesions (75.2%) were located in the colon.

Comparisons between non-cancerous lesions and ECCs are shown in Table 2. It revealed statistically significant differences in lesion length (P < 0.001) and lesion location (P < 0.001) between non-cancerous lesions and ECCs. There were no statistically significant differences in gender distribution (P = 0.332) or age distribution (P = 0.261).

Table 2 Comparison between non-cancerous and early colorectal cancer in the training set, n (%).
Item
Non-cancerous lesions
Early-stage cancer
Statistical value
P value
Sample size275129--
Age (years, mean ± SD)60.82 ± 10.8561.91 ± 9.60t = -0.9720.332
Lesion length (cm, mean ± SD)0.88 ± 0.461.95 ± 0.92t = -15.630< 0.001
Genderχ2 = 1.2640.261
    Female101 (36.7)40 (31.0)
    Male174 (63.3)89 (69.0)
Lesion locationχ2 = 15.817< 0.001
    Colon248 (90.2)97 (75.2)
    Rectum27 (9.8)32 (24.8)

Incorporating lesion length and location into a multivariate logistic regression analysis indicated that lesion length [odds ratio (OR) = 11.377, 95% confidence interval (CI): 6.839-18.928, P < 0.001] and location in rectum (OR = 2.407, 95%CI: 1.190-4.868, P = 0.015) were identified as risk factors for ECCs. The parameters are shown in Supplementary Table 1. The logistic model achieved an accuracy of 80.4%. The ROC curve is shown in Supplementary Figure 1, with an AUROC of 0.899 (95%CI: 0.868-0.930).

Variable screening from image data

The 550 non-cancerous lesion images and 258 ECC images from the training set, along with the 242 non-cancerous lesion images and 102 ECC images from the validation set, were imported into the PyTorch. CNN training was conducted using the VGG19, ResNet50, and InceptionV3 models, respectively. All the included images underwent uniform preprocessing and standardization, converted to PyTorch tensors and perform tensor standardization. The confusion matrices of the three models in the training set are shown in Figure 2. The sensitivity, specificity, accuracy and AUROC calculated based on the confusion matrices are presented in Table 3.

Figure 2
Figure 2 Confusion matrices (top) and receiver operating characteristic curves (bottom) for the VGG19, ResNet50, and InceptionV3 models in the training set. A: Confusion matrix of the VGG19 model; B: Confusion matrix of the ResNet50 model; C: Confusion matrix of the InceptionV3 model; D: Receiver operating characteristic (ROC) curve of the VGG19 model; E: ROC curve of the ResNet50 model; F: ROC curve of the InceptionV3 model. ROC: Receiver operating characteristic.
Table 3 Test performance of VGG19 model, ResNet50 model and InceptionV3 model in the training set.
Model
Sensitivity
Specificity
Accuracy
AUROC
VGG190.780.930.880.92
ResNet500.570.940.820.84
InceptionV30.670.990.890.95
Ensemble deep learning prediction0.850.930.900.89

All three individual models showed high specificity (0.93-0.99) but relatively low sensitivity (0.57-0.78) in the training set. Therefore, the final decision rule for deep learning was established as follows: If at least one of the three models identified a lesion as ECC, the output was classified as ECC; otherwise, it was classified as a non-cancerous lesion. This ensemble deep learning prediction achieved a sensitivity of 0.85, specificity of 0.93, accuracy of 0.90, and an AUC of 0.89 in the training set, indicating excellent diagnostic performance.

To visually illustrate the inference process of the three models, Figure 3 displays their Gradient-weighted Class Activation Mapping (Grad-CAM) results. The left side are the original images. while the right side shows the weight maps generated by the Grad-CAM algorithm for the VGG19 model, ResNet50 model, and InceptionV3 model, respectively. Regions are color-coded from blue to red, with redder areas representing greater importance to the model’s decision-making process.

Figure 3
Figure 3 The Gradient-weighted Class Activation Mapping algorithms for three models. The right side shows the weight maps generated by the Gradient-weighted Class Activation Mapping algorithm for the VGG19, ResNet50, and InceptionV3 models, respectively. The color scale ranges from blue to red, with redder regions indicating greater importance to the model’s decision.
Establishment and validation of the deep learning-based multimodal model

Based on above outcomes, lesion length, lesion location, and the ensemble deep learning prediction were selected as input variables. A nomogram for discriminating between non-cancerous lesions and ECC was subsequently developed using the forward likelihood ratio method for variable selection. The resulting nomogram is presented in Figure 4. The nomogram-derived probability was dichotomized using a threshold of 0.5, such that cases with a predicted probability greater than 0.5 were classified as ECC, while those with a probability of 0.5 or lower were classified as non-cancerous lesions.

Figure 4
Figure 4 Nomogram of the multimodal model for distinguishing non-cancerous lesions from early colorectal cancers. Deep learning prediction represents the output of the majority-vote ensemble of VGG19, ResNet50, and InceptionV3 (early colorectal cancer if at least 1 model positive).

To validate the diagnostic efficacy of the multimodal model and compare its performance against traditional models, the parameters obtained from the VGG19 model, ResNet50 model, InceptionV3 model, logistic regression model (referred to as the clinical parameter group in the chart), the ensemble deep learning prediction, and multimodal model parameters were used to judge the validation set's clinical data and endoscopic images. Concurrently, three junior physicians with less than three years of endoscopic experience (the junior physician group) and three senior physicians with over three years of endoscopic experience (the senior physician group) diagnosis the validation set. The ROC curves result is shown in Figure 5 below. The sensitivity, specificity, accuracy and the AUROC of each group are summarized in Table 4.

Figure 5
Figure 5 Receiver operating characteristic curves of each group of models in the validation set. ROC: Receiver operating characteristic.
Table 4 Diagnostic performance of the multimodal model, junior doctors, and senior doctors in the validation set.
Regimentation
Sensitivity
Specificity
Accuracy
AUROC
The VGG19 model0.430.820.700.75
The ResNet50 model0.350.940.760.86
The InceptionV3 model0.710.880.830.88
Clinical parameter group0.530.940.810.88
Ensemble deep learning prediction0.750.880.840.82
    Junior endoscopist group0.760.730.720.77
    Senior endoscopist group0.780.850.830.82
    Multimodal model group0.800.940.900.95

Within the validation set, the multimodal model achieved a sensitivity, specificity, accuracy was 0.80, 0.94, 0.90 respectively, and AUROC was 0.95 (95%CI: 0.91-0.99). The combination of the ensemble deep learning prediction and clinical parameters effectively enhances the sensitivity of the model, it improves the model's ability to identify early cancer lesions. Compared to both the junior and senior endoscopist groups, the multimodal model demonstrated superior specificity and accuracy, while showing slightly better sensitivity.

To validate the consistency between the multimodal model's predictions and actual results, Supplementary Figure 2 shows the calibration curves of the multimodal model in the training and validation sets. The ideal curve (long dashed line), calibration curve (short dashed line), and predicted curve (solid line) are close, indicating good agreement between the model’s predictions and actual results. To assess the net clinical benefit of the multimodal model, Supplementary Figure 3 shows the decision curves from the calibration curves of the multimodal model in the training and validation sets. The red line representing the model's decision acceptance threshold lies above both the full acceptance curve and the full rejection curve, indicating that accepting the model’s decisions yields a certain clinical net benefit.

DISCUSSION

Colorectal cancer has four principal pathways: (1) Adenoma-adenocarcinoma, (2) Serrated lesions- carcinoma; (3) Inflammatory bowel disease related cancer; and (4) colorectal de novo carcinoma[15-18]. Among these pathways, The “normal mucosa-adenoma-adenocarcinoma” pathway accounts for 70%-80% of sporadic colorectal cancer[19]. The detection and resection of precancerous lesions and ECC effectively interrupt this progression, thereby preventing the development or advancement of colorectal malignancy[20,21].

ECC refers to malignant tumors confined to the mucosal and submucosal layers. With the development of endoscopic treatment techniques, patients who accept endoscopic resection with negative margins, no lymph node metastasis, and no vascular metastasis may avoid additional surgical intervention, effectively improving patient prognosis[22,23]. ECC remains challenging to detect. To address this, we developed a deep learningbased multimodal model that combines lesion length, location, and the ensemble deep learning prediction. The model demonstrated favorable sensitivity, specificity, and accuracy in the validation set.

In colorectal lesions, previous studies have reported higher incidence rates of malignant tumors in the left colon and rectum compared to the right colon. Lee et al[24], through a systematic review including 6 meta-analyses, 12 reviews, 62 observational studies, and 7 supplementary literatures, suggested that the left colon and rectum are more prone to microsatellite instability, chromosomal instability, and epigenetic regulation abnormalities (such as DNA methylation) compared to the right colon, making them more susceptible to carcinogenesis. In this study, rectal lesions were more likely to undergo malignant transformation than colonic lesions. In addition to the aforementioned epigenetic and immunological factors, this study observed that patients with rectal lesions are more prone to symptoms such as diarrhea and constipation which lead to intestinal mucosal damage and microbiota dysbiosis may also contribute to the risk of ECC development[25-27]. Furthermore, increased lesion length is also accompanied by increased cellular heterogeneity and greater depth of invasion, elevating the risk of malignancy[28].

A logistic regression model incorporating lesion location and lesion length achieved approximately 80% accuracy in identifying ECC. Deep learning represented by CNN have favorable applications in image recognition and classification[29,30]. This study introduced three models: VGG19, ResNet50, and InceptionV3. The accuracy of these three models in distinguishing non-cancerous lesions from ECC ranged from 71% to 83%, all showing high specificity but relatively low sensitivity. However, due to differences in the algorithmic mechanisms of the three models, combining their prediction scores could complementarily improve specificity. Importantly, the introduction of the ensemble deep learning prediction also enhanced the classification capability of the traditional clinical model for ECC.

This study established a multimodal model for distinguishing non-cancerous from ECC. CNN has already been applied to distinguish lesions at different stages of the “normal mucosa-adenoma-adenocarcinoma” process. For example, Choi et al[14] used a total of 3400 images comprise normal mucosa, low-grade intraepithelial neoplasia, high-grade intraepithelial neoplasia, and adenocarcinoma developed a diagnostic system based on InceptionV3, ResNet50, and DenseNet161 models achieving a sensitivity of 77.25% and specificity of 92.42%. In our study, the integration of CNN parameters with clinical data enhanced the stability of the developed model. Cheng et al[31] developed a diagnostic model for distinguishing non-neoplastic polyps from neoplastic polyps, achieving a diagnostic accuracy of 84.49% which significantly improved the diagnostic accuracy of endoscopists (from 68.84% to 77.49%). The model we developed in this study demonstrated a significantly higher classification rate for early-stage cancer compared to endoscopists, indicating substantial potential for application in assisting endoscopists’ diagnoses.

Several limitations should be acknowledged. First, manual lesion bounding-box annotations were unavailable; thus, random cropping (60%-100%, center-biased) was applied globally without constraining the crop window to contain the lesion, posing a theoretical risk of excluding peripheral lesions. Nevertheless, consistent Grad-CAM heatmaps confirmed that all three CNN models accurately localized the lesions, indicating robust feature learning despite this augmentation risk. Second, our dataset exhibited inherent class imbalance (396 non-cancerous vs 180 ECCs). We used standard unweighted cross-entropy loss without oversampling or cost-sensitive strategies, which likely biased standalone CNNs toward the majority class, explaining their high specificity (0.92-0.98) but modest sensitivity (0.56-0.78). Encouragingly, the ensemble strategy and integration with clinical variables effectively counteracted these biases, yielding a well-balanced validation performance (sensitivity 0.80, specificity 0.94). Future prospective studies with larger cohorts, precise lesion annotations, and advanced imbalance techniques (e.g., focal loss or SMOTE) are warranted to further enhance model robustness.

CONCLUSION

In conclusion, this study based on clinical data and white-light endoscopic images developed a multimodal model to distinguish non-cancerous lesions from ECC. The model demonstrates good diagnostic performance and may improve the diagnostic accuracy of ECC for endoscopists.

References
1.  Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71:209-249.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 76817]  [Cited by in RCA: 71791]  [Article Influence: 14358.2]  [Reference Citation Analysis (111)]
2.  Han B, Zheng R, Zeng H, Wang S, Sun K, Chen R, Li L, Wei W, He J. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent. 2024;4:47-53.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1395]  [Cited by in RCA: 1756]  [Article Influence: 878.0]  [Reference Citation Analysis (10)]
3.  Zauber AG, Winawer SJ, O'Brien MJ, Lansdorp-Vogelaar I, van Ballegooijen M, Hankey BF, Shi W, Bond JH, Schapiro M, Panish JF, Stewart ET, Waye JD. Colonoscopic polypectomy and long-term prevention of colorectal-cancer deaths. N Engl J Med. 2012;366:687-696.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 2583]  [Cited by in RCA: 2469]  [Article Influence: 176.4]  [Reference Citation Analysis (18)]
4.  Hayashi N, Tanaka S, Hewett DG, Kaltenbach TR, Sano Y, Ponchon T, Saunders BP, Rex DK, Soetikno RM. Endoscopic prediction of deep submucosal invasive carcinoma: validation of the narrow-band imaging international colorectal endoscopic (NICE) classification. Gastrointest Endosc. 2013;78:625-632.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 386]  [Cited by in RCA: 329]  [Article Influence: 25.3]  [Reference Citation Analysis (9)]
5.  Zhang Y, Chen HY, Zhou XL, Pan WS, Zhou XX, Pan HH. Diagnostic efficacy of the Japan Narrow-band-imaging Expert Team and Pit pattern classifications for colorectal lesions: A meta-analysis. World J Gastroenterol. 2020;26:6279-6294.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in CrossRef: 14]  [Cited by in RCA: 15]  [Article Influence: 2.5]  [Reference Citation Analysis (5)]
6.  IJspeert JE, Bastiaansen BA, van Leerdam ME, Meijer GA, van Eeden S, Sanduleanu S, Schoon EJ, Bisseling TM, Spaander MC, van Lelyveld N, Bargeman M, Wang J, Dekker E; Dutch Workgroup serrAted polypS & Polyposis (WASP). Development and validation of the WASP classification system for optical diagnosis of adenomas, hyperplastic polyps and sessile serrated adenomas/polyps. Gut. 2016;65:963-970.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 218]  [Cited by in RCA: 189]  [Article Influence: 18.9]  [Reference Citation Analysis (0)]
7.  Pimentel-Nunes P, Libânio D, Bastiaansen BAJ, Bhandari P, Bisschops R, Bourke MJ, Esposito G, Lemmers A, Maselli R, Messmann H, Pech O, Pioche M, Vieth M, Weusten BLAM, van Hooft JE, Deprez PH, Dinis-Ribeiro M. Endoscopic submucosal dissection for superficial gastrointestinal lesions: European Society of Gastrointestinal Endoscopy (ESGE) Guideline - Update 2022. Endoscopy. 2022;54:591-622.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 617]  [Cited by in RCA: 634]  [Article Influence: 158.5]  [Reference Citation Analysis (4)]
8.  Sharma P, Frye J, Frizelle F. Accuracy of visual prediction of pathology of colorectal polyps: how accurate are we? ANZ J Surg. 2014;84:365-370.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 6]  [Cited by in RCA: 7]  [Article Influence: 0.6]  [Reference Citation Analysis (0)]
9.  Ladabaum U, Fioritto A, Mitani A, Desai M, Kim JP, Rex DK, Imperiale T, Gunaratnam N. Real-time optical biopsy of colon polyps with narrow band imaging in community practice does not yet meet key thresholds for clinical decisions. Gastroenterology. 2013;144:81-91.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 186]  [Cited by in RCA: 189]  [Article Influence: 14.5]  [Reference Citation Analysis (9)]
10.  Mitsala A, Tsalikidis C, Pitiakoudis M, Simopoulos C, Tsaroucha AK. Artificial Intelligence in Colorectal Cancer Screening, Diagnosis and Treatment. A New Era. Curr Oncol. 2021;28:1581-1607.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 233]  [Cited by in RCA: 169]  [Article Influence: 33.8]  [Reference Citation Analysis (15)]
11.  Rompianesi G, Pegoraro F, Ceresa CD, Montalti R, Troisi RI. Artificial intelligence in the diagnosis and management of colorectal cancer liver metastases. World J Gastroenterol. 2022;28:108-122.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in CrossRef: 127]  [Cited by in RCA: 105]  [Article Influence: 26.3]  [Reference Citation Analysis (2)]
12.  Zhou D, Tian F, Tian X, Sun L, Huang X, Zhao F, Zhou N, Chen Z, Zhang Q, Yang M, Yang Y, Guo X, Li Z, Liu J, Wang J, Wang J, Wang B, Zhang G, Sun B, Zhang W, Kong D, Chen K, Li X. Diagnostic evaluation of a deep learning model for optical diagnosis of colorectal cancer. Nat Commun. 2020;11:2961.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 88]  [Cited by in RCA: 68]  [Article Influence: 11.3]  [Reference Citation Analysis (0)]
13.  Kader R, Cid-Mejias A, Brandao P, Islam S, Hebbar S, Puyal JG, Ahmad OF, Hussein M, Toth D, Mountney P, Seward E, Vega R, Stoyanov D, Lovat LB. Polyp characterization using deep learning and a publicly accessible polyp video database. Dig Endosc. 2023;35:645-655.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 15]  [Cited by in RCA: 17]  [Article Influence: 5.7]  [Reference Citation Analysis (1)]
14.  Choi SJ, Kim ES, Choi K. Prediction of the histology of colorectal neoplasm in white light colonoscopic images using deep learning algorithms. Sci Rep. 2021;11:5311.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 3]  [Cited by in RCA: 15]  [Article Influence: 3.0]  [Reference Citation Analysis (0)]
15.  Atkin W, Wooldrage K, Brenner A, Martin J, Shah U, Perera S, Lucas F, Brown JP, Kralj-Hans I, Greliak P, Pack K, Wood J, Thomson A, Veitch A, Duffy SW, Cross AJ. Adenoma surveillance and colorectal cancer incidence: a retrospective, multicentre, cohort study. Lancet Oncol. 2017;18:823-834.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 209]  [Cited by in RCA: 204]  [Article Influence: 22.7]  [Reference Citation Analysis (4)]
16.  Murakami T, Kurosawa T, Fukushima H, Shibuya T, Yao T, Nagahara A. Sessile serrated lesions: Clinicopathological characteristics, endoscopic diagnosis, and management. Dig Endosc. 2022;34:1096-1109.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 40]  [Cited by in RCA: 40]  [Article Influence: 10.0]  [Reference Citation Analysis (2)]
17.  Mueller JD, Bethke B, Stolte M. Colorectal de novo carcinoma: a review of its diagnosis, histopathology, molecular biology, and clinical relevance. Virchows Arch. 2002;440:453-460.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 19]  [Cited by in RCA: 22]  [Article Influence: 0.9]  [Reference Citation Analysis (1)]
18.  Gu Y, Zhao H, Zheng L, Zhou C, Han Y, Wu A, Jia Z, Xia T, Zhi Q. Non-coding RNAs and colitis-associated cancer: Mechanisms and clinical applications. Clin Transl Med. 2023;13:e1253.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 12]  [Cited by in RCA: 15]  [Article Influence: 5.0]  [Reference Citation Analysis (0)]
19.  Elsayed I, Li L, Sheahan K, Moran B, Bakheit S, Wang X. Adenoma to carcinoma: A portrait of molecular and immunological profiles of colorectal sporadic tumors. Int Immunopharmacol. 2021;100:108168.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 7]  [Reference Citation Analysis (0)]
20.  Patel SG, Karlitz JJ, Yen T, Lieu CH, Boland CR. The rising tide of early-onset colorectal cancer: a comprehensive review of epidemiology, clinical features, biology, risk factors, prevention, and early detection. Lancet Gastroenterol Hepatol. 2022;7:262-274.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 664]  [Cited by in RCA: 622]  [Article Influence: 155.5]  [Reference Citation Analysis (16)]
21.  Eng C, Jácome AA, Agarwal R, Hayat MH, Byndloss MX, Holowatyj AN, Bailey C, Lieu CH. A comprehensive framework for early-onset colorectal cancer research. Lancet Oncol. 2022;23:e116-e128.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 192]  [Cited by in RCA: 182]  [Article Influence: 45.5]  [Reference Citation Analysis (9)]
22.  Tonini V, Zanni M. Why is early detection of colon cancer still not possible in 2023? World J Gastroenterol. 2024;30:211-224.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in CrossRef: 36]  [Cited by in RCA: 34]  [Article Influence: 17.0]  [Reference Citation Analysis (9)]
23.  Wu Y, Jiao N, Zhu R, Zhang Y, Wu D, Wang AJ, Fang S, Tao L, Li Y, Cheng S, He X, Lan P, Tian C, Liu NN, Zhu L. Identification of microbial markers across populations in early detection of colorectal cancer. Nat Commun. 2021;12:3063.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 202]  [Cited by in RCA: 180]  [Article Influence: 36.0]  [Reference Citation Analysis (6)]
24.  Lee GH, Malietzis G, Askari A, Bernardo D, Al-Hassi HO, Clark SK. Is right-sided colon cancer different to left-sided colorectal cancer? - a systematic review. Eur J Surg Oncol. 2015;41:300-308.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 341]  [Cited by in RCA: 332]  [Article Influence: 30.2]  [Reference Citation Analysis (7)]
25.  Liang L, Kong C, Li J, Liu G, Wei J, Wang G, Wang Q, Yang Y, Shi D, Li X, Ma Y. Distinct microbes, metabolites, and the host genome define the multi-omics profiles in right-sided and left-sided colon cancer. Microbiome. 2024;12:274.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 29]  [Cited by in RCA: 32]  [Article Influence: 16.0]  [Reference Citation Analysis (0)]
26.  Baran B, Mert Ozupek N, Yerli Tetik N, Acar E, Bekcioglu O, Baskin Y. Difference Between Left-Sided and Right-Sided Colorectal Cancer: A Focused Review of Literature. Gastroenterology Res. 2018;11:264-273.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 439]  [Cited by in RCA: 401]  [Article Influence: 50.1]  [Reference Citation Analysis (10)]
27.  Yang SY, Cho MS, Kim NK. Difference between right-sided and left-sided colorectal cancers: from embryology to molecular subtype. Expert Rev Anticancer Ther. 2018;18:351-358.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 69]  [Cited by in RCA: 63]  [Article Influence: 7.9]  [Reference Citation Analysis (0)]
28.  Pooler BD, Kim DH, Matkowskyj KA, Newton MA, Halberg RB, Grady WM, Hassan C, Pickhardt PJ. Growth rates and histopathological outcomes of small (6-9 mm) colorectal polyps based on CT colonography surveillance and endoscopic removal. Gut. 2023;72:2321-2328.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 14]  [Cited by in RCA: 13]  [Article Influence: 4.3]  [Reference Citation Analysis (0)]
29.  Xu H, Tang RSY, Lam TYT, Zhao G, Lau JYW, Liu Y, Wu Q, Rong L, Xu W, Li X, Wong SH, Cai S, Wang J, Liu G, Ma T, Liang X, Mak JWY, Xu H, Yuan P, Cao T, Li F, Ye Z, Shutian Z, Sung JJY. Artificial Intelligence-Assisted Colonoscopy for Colorectal Cancer Screening: A Multicenter Randomized Controlled Trial. Clin Gastroenterol Hepatol. 2023;21:337-346.e3.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 183]  [Cited by in RCA: 159]  [Article Influence: 53.0]  [Reference Citation Analysis (12)]
30.  Waljee AK, Weinheimer-Haus EM, Abubakar A, Ngugi AK, Siwo GH, Kwakye G, Singal AG, Rao A, Saini SD, Read AJ, Baker JA, Balis U, Opio CK, Zhu J, Saleh MN. Artificial intelligence and machine learning for early detection and diagnosis of colorectal cancer in sub-Saharan Africa. Gut. 2022;71:1259-1265.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 59]  [Cited by in RCA: 31]  [Article Influence: 7.8]  [Reference Citation Analysis (2)]
31.  Cheng Y, Li L, Bi Y, Su S, Zhang B, Feng X, Wang N, Zhang W, Yao Y, Ru N, Xiang J, Sun L, Hu K, Wen F, Wang Z, Bai L, Wang X, Wang R, Lv X, Wang P, Meng F, Xiao W, Linghu E, Chai N. Computer-aided diagnosis system for optical diagnosis of colorectal polyps under white light imaging. Dig Liver Dis. 2024;56:1738-1745.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 3]  [Cited by in RCA: 4]  [Article Influence: 2.0]  [Reference Citation Analysis (0)]
Footnotes

Peer review: Externally peer reviewed.

Peer-review model: Single blind

Specialty type: Oncology

Country of origin: China

Peer-review report’s classification

Scientific quality: Grade B, Grade C

Novelty: Grade B, Grade B

Creativity or innovation: Grade B, Grade B

Scientific significance: Grade B, Grade B

P-Reviewer: Yau TO, Lecturer, PhD, United Kingdom S-Editor: Qu XL L-Editor: A P-Editor: Wang WB

Write to the Help Desk