Skip to main content
Insights into Imaging
Download PDF

A radiomics-boosted deep-learning for risk assessment of synchronous peritoneal metastasis in colorectal cancer

Insights into Imaging volume 15, Article number: 150 (2024) Cite this article

Abstract

Objectives

Synchronous colorectal cancer peritoneal metastasis (CRPM) has a poor prognosis. This study aimed to create a radiomics-boosted deep learning model by PET/CT image for risk assessment of synchronous CRPM.

Methods

A total of 220 colorectal cancer (CRC) cases were enrolled in this study. We mapped the feature maps (Radiomic feature maps (RFMs)) of radiomic features across CT and PET image patches by a 2D sliding kernel. Based on ResNet50, a radiomics-boosted deep learning model was trained using PET/CT image patches and RFMs. Besides that, we explored whether the peritumoral region contributes to the assessment of CRPM. In this study, the performance of each model was evaluated by the area under the curves (AUC).

Results

The AUCs of the radiomics-boosted deep learning model in the training, internal, external, and all validation datasets were 0.926 (95% confidence interval (CI): 0.874–0.978), 0.897 (95% CI: 0.801–0.994), 0.885 (95% CI: 0.795–0.975), and 0.889 (95% CI: 0.823–0.954), respectively. This model exhibited consistency in the calibration curve, the Delong test and IDI identified it as the most predictive model.

Conclusions

The radiomics-boosted deep learning model showed superior estimated performance in preoperative prediction of synchronous CRPM from pre-treatment PET/CT, offering potential assistance in the development of more personalized treatment methods and follow-up plans.

Critical relevance statement

The onset of synchronous colorectal CRPM is insidious, and using a radiomics-boosted deep learning model to assess the risk of CRPM before treatment can help make personalized clinical treatment decisions or choose more sensitive follow-up plans.

Key Points

  • Prognosis for patients with CRPM is bleak, and early detection poses challenges.

  • The synergy between radiomics and deep learning proves advantageous in evaluating CRPM.

  • The radiomics-boosted deep-learning model proves valuable in tailoring treatment approaches for CRC patients.

Graphical Abstract

Introduction

Colorectal cancer (CRC) ranks as the second leading cause of cancer-related deaths [1]. At the time of initial diagnosis, 5–15% of CRC patients present with peritoneal metastases [2]. Synchronous colorectal cancer peritoneal metastasis (CRPM) refers to cases where peritoneal metastasis (PM) occurs either at the time of initial diagnosis or within six months after surgery [3]. The prognosis for PM is grim, lacking standard treatment, and average life expectancy ranges from 6 to 12 months [4, 5]. Although a combination of bevacizumab and cetuximab can extend overall survival (OS) by 18.2 months for CRPM patients, its efficacy remains inferior to that for liver and lung metastases [6]. A multicenter phase III clinical trial demonstrated the potential of cytoreductive surgery (CRS) in treating CRPM, yielding a median OS of 41.2 months [7]. Accurate diagnosis of CRPM is crucial; however, the diagnostic process is challenging, and not all patients can undergo laparoscopic examination for pathological confirmation.

Noninvasive methods are commonly used to predict PM, among which serum tumor markers such as carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9) play an important role, which reflects to varying degrees of tumor invasion, proliferation, and invasion of peritoneal mesothelial cells, but not all patients exhibit elevated tumor markers, making them supplementary diagnostic tools [8, 9]. Computed tomography’s (CT) sensitivity to peritoneal lesions smaller than 0.5 cm is limited to 11–48% [10]. While diffusion-weighted magnetic resonance imaging (DW-MRI) and PET/CT are highly sensitive and specific for CRPM, they can be influenced by factors such as respiratory and intestinal motility, changes in lesion size, and pathological types. Particularly, 18F-fluorodeoxyglucose PET/CT (18F-FDG PET/CT) is valuable for diagnosing and staging various malignancies, including CRC, but distinguishing peritoneal cancer lesions is challenging due to physiological FDG uptake in the normal gut and lower uptake in certain mucinous tumors [11, 12]. Additionally, inter-reader variability in imaging diagnosis cannot be entirely avoided.

Radiomics and deep learning algorithms are gaining recognition in medical imaging analysis [13, 14]. Extracting handcrafted radiomic features allows for sensitive detection of subtle heterogeneity in tumor morphology or function. Li et al demonstrated high accuracy in predicting CRPM status by utilizing CT texture extracted from primary tumor lesions and the largest metastatic lymph node [14]. However, handcrafted radiomic features heavily rely on accurately delineating lesion boundaries. In contrast, convolutional neural networks, by directly learning specific features from input images, enhance accuracy and eliminate the need for precise lesion depiction [15].

This study leverages the complementary nature of handcrafted radiomics and deep learning to develop a radiomics-boosted deep learning model for preoperative risk assessment of synchronous CRPM based on PET/CT images. Notably, there are no reports on the application of PET/CT-based radiomics and deep learning methods for predicting synchronous CRPM.

Methods and materials

Patient selection

This retrospective study received approval from the Ethics Committee of the Affiliated Hospital of Nantong University, and informed consent was waived due to its retrospective nature. The study included 220 patients diagnosed with CRC at Nantong University Affiliated Hospital between June 2016 and August 2023. Prior to March 2023, patients were randomly divided into training (n = 123), internal validation (n = 41), and external validation cohorts at a 6:2:2 ratio. Subsequently, all patients diagnosed from March 2023 to August 2023 were assigned to external validation cohorts (n = 56). Eighty with PM and 140 non-metastasis (NM) were included. Inclusion criteria comprised: (1) undergoing FDG-PET/CT scans before any treatment; (2) confirming CRC diagnosis through surgery or biopsy; (3) availability of follow-up data and clinical-pathologic information. Exclusion criteria included: (1) undergoing neoadjuvant chemotherapy and radiotherapy before surgery; (2) poor PET/CT image quality affecting accurate labeling; (3) abdominal trauma, abdominopelvic infection, or concurrent lesions of other malignant tumors. Detailed cohort inclusion is presented in Fig. 1. Clinicopathological data and lab results were collected within two weeks before the PET/CT scan. Details of synchronous CRPM status evaluation are provided in Supplementary A1.

Fig. 1
figure 1

Participant recruitment flowchart for the study

Full size image

Region of interest (ROI) patch generation

Details of 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG-PET/CT) examinations, image preprocessing, and subjective CT finding evaluation are outlined in Supplementary A2. Using Lifex software (version 7.23) [16], original regions of interest (ROIs) were drawn along the lesion contour on the largest tumor cross-sectional image section in the axial direction. A minimum rectangle boundary was created around each manual original ROI, and cropping was performed to obtain original ROI patches. To explore the potential contributions of surrounding tissue, we expanded 10 pixels around the minimum bounding box of the ROI, creating expanded ROI patches (Fig. 2). All images were resized to 224 × 224.

Fig. 2
figure 2

Workflow of RFM calculation in the study. RFM, radiomics feature map

Full size image

Radiomic feature maps (RFMs) generation

Classical radiomics is commonly utilized to capture the overall texture of ROIs, but it may not effectively discern subtle texture variations. To address this limitation, we devised a workflow for computing RFMs, as illustrated in Fig. 2. Utilizing a 7 × 7 matrix-size kernel, we extracted 75 radiomic features from the ROI, specifically including GLCM × 24, GLDM × 14, GLRLM × 16, GLSZM × 16, and NGTDM × 5 (refer to Table 1) [17]. Employing the kernel, we generated 75 feature maps with the same dimensions as the ROI patch by sliding it across the ROI patches, with the center value of the kernel being filled into each pixel.

Table 1 Seventy-five radiomics features included in this study
Full size table

For each feature, the feature value was assigned by centering at the ROI, utilizing a moving kernel across the ROI patch as a sliding window operation. This process resulted in the formation of 75 feature maps, maintaining the same dimensions as the original ROI patches.

To mitigate feature redundancy, for each patient in the training cohort, we calculated Pearson correlation coefficients (r) for each pair of RFMs, leading to the creation of 123 covariance matrices. Highly correlated feature maps with r > 0.95 were then excluded in the subsequent average Pearson covariance matrix [18].

Neural network architecture

We explored two neural network architectures, namely ResNet34 (Fig. 3A) and ResNet50 (Fig. 3B). To tailor the models to the specific problem, we employed transfer learning and fine-tuning techniques, including: (1) initializing the convolutional bases with pre-trained weights from ImageNet and (2) treating the last two fully connected layers (FC) as free parameters for training specific tasks [19]. To prevent overfitting, a dropout was placed between the two FC layers, and the output utilized softmax activation. Detailed parameters of the deep learning model can be found in Supplementary A3.

Fig. 3
figure 3

Two deep neural networks studied. A ResNet34. B ResNet50

Full size image

For the pilot model, the sole input comprised original and expanded ROI patches. Grayscale ROI patches were broadcast to three channels to align with pre-trained neural network input shapes. In the radiomics-boosted model, a three-channel image was generated by stacking grayscale ROI patches and two RFMs as input variables. The two RFMs were selected based on the saliency map (SM) of the pilot model, indicating the importance of each pixel for the final classification. Pixels with higher intensity values in the SM were deemed more critical for neural networks to diagnose, so two RFMs were chosen whose average correlation with SM was the highest in the training dataset. For each model, we assessed stability through 50 runs. Subsequently, the model with the best average area under the curve (AUC) and its input variables were selected as the final model. The model closest to the average AUC was chosen to obtain output probabilities for all cases. Finally, a PET/CT radiomics-boosted deep-learning model based on network output probability was constructed using multivariable logistic analysis.

Model comparison

Radiomics-boosted deep-learning compared with classical radiomics model

We conducted a comparative analysis between the radiomics-boosted deep-learning model and the classical radiomics model to assess their effectiveness. The construction process of the classical radiomics model is detailed in Supplementary A4.

Radiomics-boosted deep-learning compared with clinical and subjective CT finding model

Univariate logistic regression analysis was employed to investigate the impact of clinical information and subjective CT findings on CRPM risk. The clinical and subjective CT finding model was developed using multivariable logistic analysis with backward stepwise selection.

Maximum standardized uptake value (SUVmax) based classification

SUVmax is a widely recognized semiquantitative parameter in PET/CT studies. We evaluated a logistic regression model based on SUVmax as a baseline comparison in our cohorts [20, 21].

Statistical analyses

R version 4.2.2 (http://www.r-project.org) was utilized for statistical analyses. Categorical variables were presented as counts with proportions, while continuous variables with normally distributed and non-normally distributed data were expressed as mean ± SD, median, and interquartile range, respectively. A significance level of p < 0.05 on both sides was considered statistically significant. ROC analyses were conducted using the “pROC” package. Wilcoxon signed rank tests were employed to compare results from 50 runs in the pilot model.

Results

Clinical information and subjective CT findings

Table 2 displays the demographics and subjective CT findings across all cohorts. Patients were categorized into two groups based on follow-up results: PM and NM. Significant differences were observed in tumor infiltration into the surrounding fat and the level of CA19-9 between the two groups across all three cohorts. Primary tumor location, SUVmax, and levels of CEA exhibited differences in the training and external validation cohorts. Patient age reached statistical significance only in the internal validation cohort (p < 0.05).

Table 2 Demographic and subjective CT characteristics in the three cohorts
Full size table

Predictive performance of the pilot model

Table 3 provides a summary of quantitative comparisons, including sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), AUC, and average AUC of 50 runs, comparing the pilot models and radiomics-boosted models in both the training and internal validation cohorts. In all pilot models, using expanded ROI patches as the input variable achieved higher average AUC values in the validation cohort (p < 0.001). Comparing ResNet34 and ResNet50 models, the latter achieved average AUC gains with any CT patches and used PET expand ROI patches as input variables, yielding similar results (p < 0.001) (Fig. S1 A, B, D, and E). Ultimately, two optimal models were selected from the pilot models: ResNet50 with CT and PET expand ROI patches as input variables, exhibiting higher AUC and lower standard deviation (SD), indicating a high level of robustness (0.739 ± 0.027 and 0.798 ± 0.007, respectively). The AUC, accuracy, sensitivity, specificity, PPV, and NPV of the CT model that is closest to the average AUC value in the internal validation cohort are 0.749, 0.659, 0.600, 0.692, 0.529, and 0.750, respectively, while in the PET model, they are 0.797, 0.732, 0.933, 0.615, 0.583, and 0.941, respectively.

Table 3 Performance of pilot models with different input variables and different neural networks
Full size table

RFM selection and radiomics-boosted deep-learning model construction

Figure 4A, C illustrates the average correlation heatmap of RFM generated by expanded ROI patches from CT and PET images. After excluding highly correlated features, the RFMs of CT and PET include 37 and 27 features. The average correlation values between RFMs and the SM from the two pilot models are depicted in Fig. 4B, D, highlighting the highest-performing features with green boxes. For the RFMs of CT patches, GLDM-based small dependence emphasis (SDE) (0.145) and GLDM-based small dependence high gray level emphasis (SDHGLE) (0.146) achieved the highest average correlation. Similarly, for the RFMs of PET patches, GLCM-based difference entropy (DE) (0.288) and GLDM-based large dependence low gray level emphasis (LDLGLE) (0.258) were selected. The overall average correlation value of RFMs generated from CT is lower than PET.

Fig. 4
figure 4

RFM selection. Average correlation heatmap of RFM generated by expanded ROI patches from CT (A) and PET (C) images. Panels B, D depict the average correlation values between RFM and SM of CT and PET pilot models after removing redundant features, highlighting the highest-performing features with green boxes. RFM, radiomics feature map; SM, saliency map

Full size image

Compared to the ResNet50 using only CT expand ROI patches as input, the model’s performance significantly improved after adding RFMs (AUC from 0.739 ± 0.027 to 0.866 ± 0.018, p < 0.001) (Fig. S1 C). The AUC, accuracy, sensitivity, specificity, PPV, and NPV in the internal validation cohort are 0.869, 0.780, 0.800, 0.769, 0.667, and 0.870. Unexpectedly, the performance of the radiomics-boosted model decreased in PET (AUC from 0.798 ± 0.007 to 0.796 ± 0.008, p = 0.006) (Fig. S1 F and Table 3).

After a comprehensive comparison of various models, two best models based on PET/CT images were identified: (1) ResNet50 with CT ROI expand patches + RFMs as input. (2) ResNet50 with only PET ROI expand patches as input. The radiomics-boosted deep-learning model score was obtained by multivariable logistic regression on the output probabilities of the two models. This model not only exhibits the best predictive performance in the training and internal validation cohorts but also has high generalization ability in the external validation cohort. The AUC, accuracy, sensitivity, specificity, PPV, and NPV in the internal validation cohort are 0.897 (95% confidence interval (CI): 0.801–0.994), 0.829, 0.800, 0.846, 0.750, and 0.880, respectively, 0.885 (95% CI: 0.795–0.975), 0.821, 0.700, 0.889, 0.778, and 0.842 in the external validation cohort. In all validation datasets, these values were 0.889 (95% CI: 0.823–0.954), 0.825, 0.743, 0.871, 0765, and 0.857 (Table 4). All datasets showed differences between the NM and PM groups (Fig. 5A). According to the Hosmer-Lemeshow test, the ideal curve and the predictive calibration curve are similar in both training, internal, external, and all validation datasets (p = 0.225, 0.224, 0.447, 0.696, respectively) (Fig. 5B, C).

Table 4 Comparison of the prediction performance of four models for synchronous CRPM risk
Full size table
Fig. 5
figure 5

Performance evaluation of prediction models. Distribution of the radiomics-boosted deep-learning model score in all datasets (A). Calibration curves of the radiomics-boosted deep-learning model in training (B) and validation cohorts (C). Panels D–G show AUCs of different models in all training, internal, external, and all validation datasets. AUC, area under the curves

Full size image

Model comparison

Radiomics-boosted deep-learning compared with classical radiomics model

The AUCs of the radiomics model in all cohorts are illustrated in Fig. S2 A–F. The PET/CT classical radiomics model score, calculated by multivariable logistic regression, yielded AUCs of 0.813 (95% CI: 0.736–0.890), 0.680 (95% CI: 0.504–0.855), 0.642 (95% CI: 0.473–0.810), and 0.659 (95% CI: 0.537–0.779) in the training, internal, external, and all validation datasets, respectively. The IDI and DeLong test demonstrated that the radiomics-boosted deep-learning model improved performance compared to the classical radiomics model (IDI = 0.302, p < 0.001; DeLong test, p < 0.001) (Fig. 5D–G).

Radiomics-boosted deep-learning compared with clinical and subjective CT finding model

Three clinical information and one subjective CT feature were screened using univariate and multivariate logistic regression to construct a clinical and subjective CT finding model (Table 5). The AUCs in the training, internal, external, and all validation datasets were 0.797 (95% CI: 0.717–0.877), 0.727 (95% CI: 0.567–0.887), 0.786 (95% CI: 0.657–0.915), and 0.758 (95% CI: 0.658–0.858), respectively. The performance was still lower than the radiomics-boosted deep-learning model (IDI = 0.206, p = 0.005; DeLong test, p = 0.024) (Fig. 5D–G).

Table 5 Univariate and multivariate logistic regression analysis for clinical information and subjective CT findings in the training cohort
Full size table

SUVmax assessment

The SUVmax showed limited performance, with AUCs of 0.650 (95% CI: 0.544–0.757) in the training cohort and 0.680 (95% CI: 0.506–0.853), 0.672 (95% CI: 0.518–0.826), and 0.686 (95% CI: 0.575–0.798) in the internal, external, and all validation datasets, respectively (Fig. 5D–G). Detailed data about AUC, accuracy, sensitivity, specificity, PPV, and NPV of all models are presented in Table 4.

Discussion

This study utilized a new complementary approach of radiomics and deep learning for risk prediction of synchronous CRPM. The proposed radiomics-boosted deep learning model is completely superior to the clinical and subjective CT finding model, classical radiomics model, and SUVmax assessment, achieving optimal performance. We believe that for high-risk patients, it is necessary to search for metastatic lesions in more detail or develop a more compact and sensitive follow-up plan.

The independent predictive factors for CRPM in this study include elevated levels of CEA and CA19-9, which are consistent with previous research findings [22,23,24,25]. In addition, we also found that tumors situated in the right colon have a higher propensity for PM (p = 0.020). Previous research indicates that compared to adenocarcinomas in the left colon (32.5%), mucinous adenocarcinomas (42.3%) and signet ring cell carcinomas (48.8%) are more prevalent in the right colon, with respective probabilities of PM at 20.1%, 48.2%, and 51.2% (p < 0.001) [26]. Recent studies in CRC have explored the application of radiomics and deep learning techniques [27,28,29,30,31]. Li et al [14] developed a radionics-clinical fusion model based on texture features from primary tumor lesions and the largest metastatic lymph node, achieving good performance in predicting synchronous CRPM (training set AUC: 0.855, validation set: 0.793). Yuan et al [32] employed a ResNet3D + SVM-based deep learning framework, demonstrating potential in PM detection, albeit requiring substantial computational resources. Zhang et al [33] successfully developed a PM detection model with robust generalization using meta-learning algorithms, despite limited raw data availability, achieving an AUC of 0.728.

The onset of synchronous CRPM is insidious and symptoms lack specificity. National guidelines recommend the hematological, imaging examination, diagnostic laparoscopy, and cytological examination of abdominal fluid or perfusion fluid as primary diagnostic tools for CRPM [12, 34,35,36]. Nonetheless, due to the limitations of various methods, a significant number of patients are not diagnosed with PM until surgical exploration [37]. This study, involving a minimum six-month follow-up through pathology or imaging, developed classical radiomics and radiomics-boosted deep learning model based on pre-treatment PET/CT images. These models effectively assess the current risk and predict short-term significant progression of CRPM in patients, even in those initially assessed as negative. It can remind doctors to search for metastatic lesions in more detail or develop more sensitive follow-up plans. In this study, four CT and two PET radiomics features were constructed traditional radiomics models for synchronous CRPM risk evaluation (Supplementary A5). Notably, in our study, CT-derived features emphasized image texture complexity and uniformity, whereas PET features, exclusively first-order, depicted pixel or voxel distribution within the ROI, akin to physicians’ film interpretations, ignoring spatial correlations [17]. However, the classical radiomics model, reliant on predefined features, exhibits poor prediction accuracy (AUC: 0.659 (95% CI: 0.537–0.779) across all validation datasets). Previous research has highlighted synergies between deep learning and radiomics features, prompting this study to boost model efficacy through deep learning techniques, whilst leveraging the strengths of radiomics features [38,39,40,41].

To overcome the challenge of deep learning’s extensive training dataset requirement, we applied transfer learning and data augmentation to expand our dataset and prevent overfitting. The streamlined architecture and efficient training time of ResNet effectively captured predictive features, as demonstrated in our study and supported by other research [3, 42].

To comprehend the impact of tumor-surrounding tissues on classification, we annotated additional image patches encompassing the surrounding regions of the original ROI. The inclusion of surrounding tissue significantly enhanced the performance of both the CT and PET models, resulting in an average AUC of 0.739 ± 0.027 and 0.798 ± 0.007 in the validation cohort. This observation is consistent with the subjective assessment of tumor-surrounding fat infiltration on CT images (p < 0.05 in all cohorts). In CRC, the pT4 category, indicating serosal involvement by tumor cells, has been identified as a crucial independent prognostic factor, surpassing local spread and lymph node involvement [43]. Santvoort et al [44] advocate for diagnostic laparoscopy before selective resection in patients with radiological suspicion of T4 CRC. The introduction of RFMs is a noteworthy highlight, capturing subtle texture variations within ROIs driven by anatomical factors, as opposed to radiomic features calculated as scalar values. Prior studies on lung diseases have validated the rationality of this method, suggesting that the potential functional information in RFMs can enhance the accuracy of preoperative risk assessment for PM in CRC patients [18, 45]. The superimposition of CT images and RFMs yielded significant performance benefits, with GLDM-based SDE and GLDM-based SDHGLE having the highest correlation to SM. GLDM-based SDE assesses the distribution of small dependencies, with a greater value indicating smaller dependence and less homogeneous textures. GLDM-based SDHGLE measures the joint distribution of small dependence with higher gray-level values. The tumor boundary in two sets of RFMs is better defined than in the original image, and if fat infiltration is present around the tumor in the original image, the pixel value in the corresponding position in the RFMs decreases accordingly (Fig. 6).

Fig. 6
figure 6

Radiomics-boosted deep learning model visualization. The first two rows depict patients with PM, and the last two rows of NM patients. The first and fifth columns are the original image patches of the PET/CT image. The surrounding area of the tumor in PM patients is relatively blurry on CT images. The second and third columns are the RFM generated based on the original CT slices, the relatively blurry tumor boundaries on CT patches are interrupted at the corresponding positions in RFM (GLDM SDE, GLDM SDHGLE). The fourth and sixth columns are visualizations of the radiomics-boosted deep learning model based on PET/CT images. The red color highlights the ROI. The CT model focuses on the surrounding tissues of the tumor, while the PET model focuses more on the distribution of SUVmax intensity. PM, peritoneal metastasis; NM, non-metastasis; RFM, radiomics feature map; GLDM SDE, GLDM-based small dependence emphasis; GLDM SDHGLE, GLDM-based small dependence high gray level emphasis

Full size image

The neural network model based on PET original ROI patches exhibits good predictive performance, but after radiomics enhancement, the performance is suppressed. This may be attributed to the rich tissue metabolism information in the original image, but a low number of pixels and the lack of anatomical texture information in the image. These results align with recent research, indicating a close relationship between FDG metabolism levels in primary tumor tissue and CRPM, with mucus components prone to causing PM exhibiting lower FDG uptake, although the AUC value was relatively low in this study [9, 46, 47].

Figure 6 illustrates the inference process for two deep-learning models. Pixels in SMs represent attention in the deep learning model, and hot regions are colored based on attention patterns. The model exhibits more attention to the surrounding area of the tumor for CT images, while the attention in PET images demonstrates significant differences in FDG metabolism.

Several limitations should be acknowledged in this study. Firstly, it is a retrospective and single-center study, with a relatively small sample size and potential bias, which may overestimate the performance of the model. Further studies should be conducted in a different database (e.g., multicentre, larger sample size) in the future. Secondly, the diagnosis of some CRPM patients relied on follow-up due to the inability to undergo surgery or laparoscopic exploration for confirmation, resulting in a lack of a gold standard diagnosis. Thirdly, the study utilized a 2D convolutional neural network, converting three-dimensional lesion information into two dimensions, which may disrupt the spatial topological relationship between different tumor layers, potentially leading to the loss of key information for overall tumor evaluation. Fourthly, the classical radiomics model only extracted features from tumor lesions and did not include the surrounding area of the tumor in this study. Lastly, we observed that the model based on clinical and subjective CT findings demonstrated high sensitivity. In the future, integrating the deep learning model with clinical information may achieve higher diagnostic efficiency.

In conclusion, our results indicate that the Radiomics-boosted deep-learning model surpasses the classical radiomics model, SUVmax model, and clinical and subjective CT finding model. This could potentially assist physicians in making more personalized treatment decisions and follow-up plans for patients.

Data availability

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

Abbreviations

AUC:

Area under the curve

CA19-9:

Carbohydrate antigen 19-9

CEA:

Carcinoembryonic antigen

95% CI:

95% Confidence interval

CRC:

Colorectal cancer

CRPM:

Colorectal cancer peritoneal metastasis

CT:

Computed tomography

FC:

Fully connected

18F-FDG-PET/CT:

18F-fluorodeoxyglucose positron emission tomography/computed tomography

NM:

Non-metastasis

NPV:

Negative predictive value

PM:

Peritoneal metastasis

PPV:

Positive predictive value

RFM:

Radiomic feature map

ROI:

Region of interest

SD:

Standard deviation

SM:

Significance map

SUVmax:

Maximum standardized uptake value

References

  1. Keum N, Giovannucci E (2019) Global burden of colorectal cancer: emerging trends, risk factors and prevention strategies. Nat Rev Gastroenterol Hepatol 16:713–732

    Article  PubMed  Google Scholar 

  2. Cervantes A, Adam R, Rosello S et al (2023) Metastatic colorectal cancer: ESMO clinical practice guideline for diagnosis, treatment and follow-up. Ann Oncol 34:10–32

    Article  CAS  PubMed  Google Scholar 

  3. Kitamura G, Chung CY, Moore 2nd BE (2019) Ankle fracture detection utilizing a convolutional neural network ensemble implemented with a small sample, de novo training, and multiview incorporation. J Digit Imaging 32:672–677

    Article  PubMed  PubMed Central  Google Scholar 

  4. Heuvelings DJI, Wintjens A, Moonen L et al (2023) Predictive genetic biomarkers for the development of peritoneal metastases in colorectal cancer. Int J Mol Sci 24:12830

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Knotts CM, Osman MA, Aderonmu AA et al (2023) Defining the values and quality of life of cancer survivors following cytoreductive surgery and hyperthermic intraperitoneal chemotherapy: an international survey study. Ann Surg Oncol 30:7825–7832

    Article  PubMed  Google Scholar 

  6. Klaver YL, Leenders BJ, Creemers GJ et al (2013) Addition of biological therapies to palliative chemotherapy prolongs survival in patients with peritoneal carcinomatosis of colorectal origin. Am J Clin Oncol 36:157–161

    Article  CAS  PubMed  Google Scholar 

  7. Quenet F, Elias D, Roca L et al (2021) Cytoreductive surgery plus hyperthermic intraperitoneal chemotherapy versus cytoreductive surgery alone for colorectal peritoneal metastases (PRODIGE 7): a multicentre, randomised, open-label, phase 3 trial. Lancet Oncol 22:256–266

    Article  CAS  PubMed  Google Scholar 

  8. Yao L, Shao H, Zhang X, Huang X (2023) A novel risk model for predicting peritoneal metastasis in colorectal cancer based on the SEER database. J Cancer Res Clin Oncol 149:15989–16000

    Article  PubMed  Google Scholar 

  9. Chen B, Zhou J, Ma Y, Sun Q, Ren J, Wang D (2023) Evaluation of multiple biological indicators for the combined diagnosis of metastases from colorectal cancer-a retrospective study based on 1163 patients. World J Surg Oncol 21:229

    Article  PubMed  PubMed Central  Google Scholar 

  10. Kim DH, Moreno CC, Pickhardt PJ (2018) Computed tomography colonography: pearls and pitfalls. Radiol Clin North Am 56:719–735

    Article  PubMed  Google Scholar 

  11. Elekonawo FMK, Starremans B, Laurens ST et al (2020) Can [(18)F]F-FDG PET/CT be used to assess the pre-operative extent of peritoneal carcinomatosis in patients with colorectal cancer? Abdom Radiol (NY) 45:301–306

    Article  PubMed  Google Scholar 

  12. van ‘t Sant I, Engbersen MP, Bhairosing PA et al (2020) Diagnostic performance of imaging for the detection of peritoneal metastases: a meta-analysis. Eur Radiol 30:3101–3112

    Article  PubMed  Google Scholar 

  13. Song Z, Liu T, Shi L et al (2021) The deep learning model combining CT image and clinicopathological information for predicting ALK fusion status and response to ALK-TKI therapy in non-small cell lung cancer patients. Eur J Nucl Med Mol Imaging 48:361–371

    Article  CAS  PubMed  Google Scholar 

  14. Li M, Sun K, Dai W et al (2020) Preoperative prediction of peritoneal metastasis in colorectal cancer using a clinical-radiomics model. Eur J Radiol 132:109326

    Article  PubMed  Google Scholar 

  15. Castiglioni I, Rundo L, Codari M et al (2021) AI applications to medical images: from machine learning to deep learning. Phys Med 83:9–24

    Article  PubMed  Google Scholar 

  16. Nioche C, Orlhac F, Boughdad S et al (2018) LIFEx: a freeware for radiomic feature calculation in multimodality imaging to accelerate advances in the characterization of tumor heterogeneity. Cancer Res 78:4786–4789

    Article  CAS  PubMed  Google Scholar 

  17. van Griethuysen JJM, Fedorov A, Parmar C et al (2017) Computational radiomics system to decode the radiographic phenotype. Cancer Res 77:e104–e107

    Article  PubMed  PubMed Central  Google Scholar 

  18. Yang Z, Lafata KJ, Chen X et al (2022) Quantification of lung function on CT images based on pulmonary radiomic filtering. Med Phys 49:7278–7286

    Article  CAS  PubMed  Google Scholar 

  19. Deng J, Dong W, Socher R, Li LJ, Kai L, Li F-F (2009) ImageNet: a large-scale hierarchical image database. 2009 IEEE conference on computer vision and pattern recognition 2009, Florida 20–25 June 2009

  20. Kim S, Lee JH, Park EJ et al (2023) Prediction of microsatellite instability in colorectal cancer using a machine learning model based on PET/CT radiomics. Yonsei Med J 64:320–326

    Article  PubMed  PubMed Central  Google Scholar 

  21. Li J, Yang Z, Xin B et al (2021) Quantitative prediction of microsatellite instability in colorectal cancer with preoperative PET/CT-based radiomics. Front Oncol 11:702055

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Takakura Y, Ikeda S, Imaoka Y, Urushihara T, Itamoto T (2015) An elevated preoperative serum carbohydrate antigen 19-9 level is a significant predictor for peritoneal dissemination and poor survival in colorectal cancer. Colorectal Dis 17:417–425

    Article  CAS  PubMed  Google Scholar 

  23. Shin JK, Kim HC, Lee WY et al (2019) High preoperative serum CA 19-9 levels can predict poor oncologic outcomes in colorectal cancer patients on propensity score analysis. Ann Surg Treat Res 96:107–115

    Article  CAS  PubMed  Google Scholar 

  24. Luo H, Shen K, Li B, Li R, Wang Z, Xie Z (2020) Clinical significance and diagnostic value of serum NSE, CEA, CA19-9, CA125 and CA242 levels in colorectal cancer. Oncol Lett 20:742–750

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Lakemeyer L, Sander S, Wittau M, Henne-Bruns D, Kornmann M, Lemke J (2021) Diagnostic and prognostic value of CEA and CA19-9 in colorectal cancer. Diseases 9:21

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  26. Hugen N, van de Velde CJH, de Wilt JHW, Nagtegaal ID (2014) Metastatic pattern in colorectal cancer is strongly influenced by histological subtype. Ann Oncol 25:651–657

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  27. Ying M, Pan J, Lu G et al (2022) Development and validation of a radiomics-based nomogram for the preoperative prediction of microsatellite instability in colorectal cancer. BMC Cancer 22:524

    Article  PubMed  PubMed Central  Google Scholar 

  28. Wang Q, Xu J, Wang A et al (2023) Systematic review of machine learning-based radiomics approach for predicting microsatellite instability status in colorectal cancer. Radiol Med 128:136–148

    Article  PubMed  PubMed Central  Google Scholar 

  29. Pei Q, Yi X, Chen C et al (2022) Pre-treatment CT-based radiomics nomogram for predicting microsatellite instability status in colorectal cancer. Eur Radiol 32:714–724

    Article  CAS  PubMed  Google Scholar 

  30. Li M, Zhang J, Dan Y et al (2020) A clinical-radiomics nomogram for the preoperative prediction of lymph node metastasis in colorectal cancer. J Transl Med 18:46

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  31. Chen X, He L, Li Q et al (2023) Non-invasive prediction of microsatellite instability in colorectal cancer by a genetic algorithm-enhanced artificial neural network-based CT radiomics signature. Eur Radiol 33:11–22

    Article  PubMed  Google Scholar 

  32. Yuan Z, Xu T, Cai J et al (2022) Development and validation of an image-based deep learning algorithm for detection of synchronous peritoneal carcinomatosis in colorectal cancer. Ann Surg 275:e645–e651

    Article  PubMed  Google Scholar 

  33. Zhang H, Zhu X, Li B et al (2022) Development and validation of a meta-learning-based multi-modal deep learning algorithm for detection of peritoneal metastasis. Int J Comput Assist Radiol Surg 17:1845–1853

    Article  PubMed  Google Scholar 

  34. Yousef A, Yousef M, Zeineddine MA et al (2024) Serum tumor markers and outcomes in patients with appendiceal adenocarcinoma. JAMA Netw Open 7:e240260

    Article  PubMed  PubMed Central  Google Scholar 

  35. Mizumoto A, Takao N, Imagami T et al (2024) Cytoreductive surgery for synchronous and metachronous colorectal peritoneal dissemination: Japanese P classification and peritoneal cancer index. Ann Gastroenterol Surg 8:88–97

    Article  PubMed  Google Scholar 

  36. Canda AE, Arslan C, Terzi C et al (2018) Treatment of intraoperatively detected peritoneal carcinomatosis of colorectal origin with cytoreductive surgery and intraperitoneal chemotherapy. World J Surg Oncol 16:70

    Article  PubMed  PubMed Central  Google Scholar 

  37. Abdel-Rahman O (2020) A real-world, population-based study of the outcomes of patients with metastatic colorectal cancer to the peritoneum treated with or without cytoreductive surgery. Int J Colorectal Dis 35:719–725

    Article  PubMed  Google Scholar 

  38. Wu JM-T, Tsai M-H, Huang YZ et al (2019) Applying an ensemble convolutional neural network with Savitzky–Golay filter to construct a phonocardiogram prediction model. Applied Soft Computing 78:29–40

    Article  Google Scholar 

  39. Dong D, Fang MJ, Tang L et al (2020) Deep learning radiomic nomogram can predict the number of lymph node metastasis in locally advanced gastric cancer: an international multicenter study. Ann Oncol 31:912–920

    Article  CAS  PubMed  Google Scholar 

  40. Forghani R, Savadjiev P, Chatterjee A, Muthukrishnan N, Reinhold C, Forghani B (2019) Radiomics and artificial intelligence for biomarker and prediction model development in oncology. Comput Struct Biotechnol J 17:995–1008

    Article  PubMed  PubMed Central  Google Scholar 

  41. Li C, Dong D, Li L et al (2020) Classification of severe and critical covid-19 using deep learning and radiomics. IEEE J Biomed Health Inform 24:3585–3594

    Article  PubMed  Google Scholar 

  42. Chen T, Liu S, Li Y et al (2019) Developed and validated a prognostic nomogram for recurrence-free survival after complete surgical resection of local primary gastrointestinal stromal tumors based on deep learning. EBioMedicine 39:272–279

    Article  PubMed  Google Scholar 

  43. Zwanenburg ES, Wisselink DD, Klaver CEL et al (2022) The measured distance between tumor cells and the peritoneal surface predicts the risk of peritoneal metastases and offers an objective means to differentiate between pT3 and pT4a colon cancer. Mod Pathol 35:1991–2001

    Article  CAS  PubMed  Google Scholar 

  44. van Santvoort HC, Braam HJ, Spekreijse KR et al (2014) Peritoneal carcinomatosis in t4 colorectal cancer: occurrence and risk factors. Ann Surg Oncol 21:1686–1691

    Article  PubMed  Google Scholar 

  45. Hu Z, Yang Z, Lafata KJ, Yin FF, Wang C (2022) A radiomics-boosted deep-learning model for COVID-19 and non-COVID-19 pneumonia classification using chest x-ray images. Med Phys 49:3213–3222

    Article  CAS  PubMed  Google Scholar 

  46. Wang W, Tan GHC, Chia CS, Skanthakumar T, Soo KC, Teo MCC (2018) Are positron emission tomography-computed tomography (PET-CT) scans useful in preoperative assessment of patients with peritoneal disease before cytoreductive surgery (CRS) and hyperthermic intraperitoneal chemotherapy (HIPEC)? Int J Hyperthermia 34:524–531

    Article  PubMed  Google Scholar 

  47. Cistaro A, Cucinotta M, Cassalia L et al (2016) (18)F-FDG PET/CT, cytoreductive surgery and intraperitoneal chemohyperthermia for the therapeutic management in peritoneal carcinomatosis: a pilot study. Rev Esp Med Nucl Imagen Mol 35:232–237

    CAS  PubMed  Google Scholar 

Download references

Funding

This study has received funding from Postgraduate Research & Practice Innovation Program of Jiangsu Province (SJCX23_1808), Science and Technology Project of Nantong City (MS22022063, MS22022044), Nantong “14th Five-Year Plan” Science and Education Health Enhancement Project Young Medical Key Talents Project, Jiangsu Provincial Research Hospital (YJXYY202204-YSB18).

Author information

Author notes

  1. Ding Zhang and BingShu Zheng contributed equally to this work.

Authors and Affiliations

  1. Medical School of Nantong University, Nantong, JiangSu, China

    Ding Zhang, BingShu Zheng, LiuWei Xu & YiCong Wu

  2. Department of Nuclear Medicine, Affiliated Hospital of Nantong University, Nantong, JiangSu, China

    Ding Zhang, ShanLei Bao, ZhongHua Tan & ChunFeng Sun

  3. Department of General Surgery, Affiliated Hospital of Nantong University, Nantong, JiangSu, China

    Chen Shen

Authors
  1. Ding Zhang
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. BingShu Zheng
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. LiuWei Xu
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. YiCong Wu
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Chen Shen
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. ShanLei Bao
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. ZhongHua Tan
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. ChunFeng Sun
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

Z.D. and Z.B.S.: data analysis and manuscript writing. X.L.W. and W.Y.C.: acquisition of data. B.S.L.: statistical analysis. S.C.F., S.C., and T.Z.H.: conception and supervision. All authors read and approved the final manuscript.

Corresponding authors

Correspondence to ZhongHua Tan or ChunFeng Sun.

Ethics declarations

Ethics approval and consent to participate

Our institutional review board approved this retrospective study, and so, the requirement for informed consent was waived.

Consent for publication

The requirement for informed consent was waived.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Zhang, D., Zheng, B., Xu, L. et al. A radiomics-boosted deep-learning for risk assessment of synchronous peritoneal metastasis in colorectal cancer. Insights Imaging 15, 150 (2024). https://doi.org/10.1186/s13244-024-01733-5

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13244-024-01733-5

Keywords