A radiomics nomogram model for predicting prognosis of pancreatic ductal adenocarcinoma after high-intensity focused ultrasound surgery

Abstract

Objective

To develop and validate a radiomics nomogram for predicting the survival of patients with pancreatic ductal adenocarcinoma (PDAC) after receiving high-intensity focused ultrasound (HIFU) treatment.

Methods

A total of 52 patients with PDAC were enrolled. To select features, the least absolute shrinkage and selection operator algorithm were applied, and the radiomics score (Rad-Score) was obtained. Radiomics model, clinics model, and radiomics nomogram model were constructed by multivariate regression analysis. The identification, calibration, and clinical application of nomogram were evaluated. Survival analysis was performed using Kaplan–Meier (K–M) method.

Results

According to conclusions made from the multivariate Cox model, Rad-Score, and tumor size were independent risk factors for OS. Compared with the clinical model and radiomics model, the combination of Rad-Score and clinicopathological factors could better predict the survival of patients. Patients were divided into high-risk and low-risk groups according to Rad-Score. K–M analysis showed that the difference between the two groups was statistically significant (p < 0.05). In addition, the radiomics nomogram model indicated better discrimination, calibration, and clinical practicability in training and validation cohorts.

Conclusion

The radiomics nomogram effectively evaluates the prognosis of patients with advanced pancreatic cancer after HIFU surgery, which could potentially improve treatment strategies and promote individualized treatment of advanced pancreatic cancer.

Keywords:

• Radiomics
• nomogram
• HIFU
• pancreatic neoplasms
• survival

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a fatal disease which increasingly led to more morbidity and is expected to become the second leading cause of cancer death in some areas. PDAC is usually diagnosed at an advanced stage and ranks last among all cancer sites in terms of prognosis outcomes [1,2]. Even in patients who have the opportunity of surgical resection, the 5-year survival rate is only about 20% [3], Chemotherapy is the most effective treatment for most patients. At present, the main strategy for treating pancreatic cancer is to combine chemotherapy based on gemcitabine. However, according to previous researches, the effective rate of first-line treatment of unresectable pancreatic cancer is only 29–31.6% [4].

In recent years, an advanced treatment called high-intensity focused ultrasound (HIFU) has developed rapidly. Due to various advantages of HIFU, such as high safety, non-invasive, definite curative effect, and less complications, HIFU has gradually been well-accepted and used in the treatment of a variety of solid tumors [5]. The basic principle of HIFU is to emit focused ultrasound in vitro to focus on the tumor in vivo. Furthermore, through the thermal effect, the temperature in the focus area instantly rises to 65–100 °C, which results in coagulative necrosis of tumor cells in the target area [6]. At the same time, several studies have confirmed that HIFU can play a synergistic and sensitizing effect with chemotherapy, which extends the survival period of patients and relieve their clinical symptoms effectively [7–9]. However, the residual tumor after HIFU is inevitable, so it is important to evaluate the response of pancreatic cancer to HIFU before operation, which is helpful in guiding the treatment strategy and improving the prognosis of patients.

Radiomics extracts a large number of quantitative features from medical images in a high-throughput way, which makes it possible to analyze the heterogeneity in tumors non-invasively. Many studies have shown that radiomics features have potential applicable values in the differential diagnosis of pancreatic cancer [10,11], pathological grade [12], and prognosis prediction [13,14]. Guo et al. [11] mentioned that CT imaging features and texture parameters help distinguish pancreatic neuroendocrine carcinoma from pancreatic ductal adenocarcinoma. Attiyeh et al. [15] illustrated that CT texture feature quantitative analysis shows that heterogeneous low density tumor is a prognostic factor for low survival rate of patients with pancreatic ductal adenocarcinoma. CT images have been used in texture analysis by many researchers, it is found that it can be used as a biomarker of pathology and prognosis of patients with invasive tumor. MRI has the advantages of multi-directional imaging, multi-parameter imaging, and excellent contrast resolution. Furthermore, more valuable radiomics features may be found. According to related reports [13,16], radiomics features based on MRI images can predict the survival and prognosis of patients with pancreatic cancer. These results are encouraging. However, currently, the radiomics studies related to the prognosis of PDAC are limited.

Therefore, in this study, we aim to construct and validate a nomogram based on radiomics features, which can be used to evaluate the prognosis of PDAC after HIFU treatment. Besides, the model can provide valuable information for clinicians and guide clinical treatment strategies.

Materials and methods

Patients

This retrospective analysis obtained ethical approval from our local institutional review committee (approval QYFYWZLL27526) and waived the requirement for written informed consent. From July 2019 to May 2021, patients who met a series of inclusion criteria were included in our study. The inclusion criteria were as follows: (1) patients with PDAC confirmed by histopathology (2) patients had no possibility of radical resection (tumor could not be removed; patients could not tolerate radical surgery or patients refused surgery) (3) HIFU surgery in our hospital (4) MRI examination was performed in our hospital before operation; the exclusion criteria were as follows: (1) patients without complete clinical data, (2) patients with obvious artifacts in MRI images, and (3) patients who were followed up for <12 months. The final study population included 52 patients who were randomly assigned to a training and validation cohort at 6:4. The basic clinical features and imaging data, such as age, sex, CA19-9, TNM stage, tumor size, local or distant metastasis, were collected from their medical records. Since HIFU ablation is considered to be a supplement to standard systemic therapy, patients continued to receive chemotherapy after HIFU surgery without changing the standard chemotherapy regimen. The survival data of patients with PDAC were obtained through clinical follow-up and telephone communication.

HIFU ablation

The operation was performed with JC high-intensity focused ultrasound tumor treatment system produced by Chongqing HAIFU Medical Technology Co., Ltd. The system was mainly composed of airborne ultrasonic positioning system, ultrasonic power transmitting device, therapeutic head motion control system, medium water treatment, and circulating cooling system. To be well-prepared to be included in the study, patients were required to abstain from gas-producing food within 3 days before treatment and catharsis 10 h before treatment. The conventional treatment position is prone position. Before operation, the surgeon used the airborne ultrasound probe combined with the three-dimensional reconstruction image to determine the location, size, the shape of the tumor, and its relationship with the adjacent tissue, so as to ensure the safety of the sound channel. Each 5 mm layer, the focus was divided into multiple continuous sections and the ultrasonic power was selected according to the blood supply of the tumor. The target tumor was treated with point-line-surface-body three-dimensional therapy. During the operation, the operator could observe the overall gray increase or mass echo enhancement of the tumor in real time, until the predetermined target area was covered. Finally, contrast-enhanced ultrasound was routinely used to evaluate the therapeutic effect. If there was no obvious perfusion in the treatment area, the treatment ended; Otherwise, continue to treat the perfusion area. The treatment was completed after the second radiography reached the expected ablation.

MRI data acquisition

All images were obtained by the radiology department of our hospital. MRI was performed with 3.0 T magnetic resonance scanner (Discovery MR 750 position GE Healthcare, Milwaukee, WI) before operation. The parameters were listed as follows: T1WI (TR/TE, 400–600/10–15 ms); FS-T2WI (TR/TE, 3500–4500/90–100 ms); section thickness 4–5 mm; section spacing 0.4–0.5 mm; matrix 512 × 256; FOV 200 × 220mm. After intravenous injection of contrast medium (Gadopentetate Dimeglumine; Bayer Schering, Berlin, Germany) 0.1 mmol/kg at a rate of 2 ml/s, enhanced scanning was performed. Arterial phase and portal venous phase images were collected at 20 and 60 s after injection, respectively.

Region-of-interest (ROI) segmentation and extraction of radiomic features

The radiomics workflow is illustrated in Figure 1. Preoperative arterial phase and portal venous phase imaging were used to extract features. All areas of interest were manually delineated by radiologists on MR which includes the largest part of the tumor and excludes large blood vessels, necrosis, bleeding, and artifacts. First, the tumor was segmented independently by a radiologist with 10 years of abdominal imaging experience. Then, a radiologist with 20 years of experience independently performed tumor segmentation on randomly chosen 25 tumors to assess repeatability between interobserver. The stable and reproducible features with an intraclass correlation coefficient (ICC) value >0.75 were remained. All radiologists were blinded to clinical and survival data. All ROIs were segmented using ITKSNAPv3.6.0 from the University of Pennsylvania (www.itksnap). A total of 1132 radiomics features were extracted from each MRI sequence using AK software (Artificial Intelligence Kit v. 3.1.0.A, GE Healthcare).

Figure 1. Workflow of the development of the radiomic nomogram.

Feature selection and radiomic nomogram development

In the initial cohort, patients were randomly divided into two groups: training cohort (n = 32) and validation cohort (n = 20), with a proportion of 6:4. The linear correlation analysis was carried out with Pearson correlation analysis. Then, the least absolute shrinkage selection operator (LASSO) Cox regression method with 10-fold cross validation was used to select the most useful predictive radiomics features. The Rad-Score of each patient were calculated by a linear combination of selected features, which were weighted by their respective coefficients. The area under the receiver-operator characteristic (ROC) was used to evaluate the prediction accuracy of Rad-Score in training and Validation cohort.

The influence of radiomics features and clinicopathological features on OS was evaluated by univariate and multivariate Cox regression analysis. Then the radiomics model was established by using the variables with p < 0.05 in multivariate Cox regression analysis. In addition to the radiomics model, we also established a clinic model and radiomics nomograph model. All models were developed in the training cohort and verified in the Validation cohort. In addition, calibration curves (actual curve and prediction curve) were drawn to evaluate the calibration of nomogram. Decision curve analysis (DCA) was used to determine the clinical applicability of the nomogram by quantifying the probability of net income in the threshold range of 0–1.

Statistical analysis

Means and standard deviation were classified as continuous variables, while frequency and proportion were classified as categorical variables. In the analysis, t-test or Mann–Whitney U test were used to analyze quantitative variables. For qualitative variables, chi-square or Fisher’s exact test were used. Cox risk regression analysis was carried out on the training cohort to determine independent predictive factors and establish the prediction model. In the training cohort and validation cohort, K–M survival analysis was used to evaluate the potential association between Rad-Score and prognosis. The time-dependent ROC curve was applied to evaluate the prognostic accuracy of different models. The calibration curve was drawn to verify the accuracy and reliability of the nomogram. In addition, DCA was drawn to evaluate the applicability of nomogram in clinical practice. SPSS 26.0 (IBM Corp., Armonk, NY, USA) and R software v3.6.4 (https://www.r-project.org/) were used for all statistical analysis. The threshold of significant difference was p < 0.05.

Result

Patient characteristics

A total of 52 patients were included in this study, including 32 males and 20 females, aged from 42 to 78 years, with an average of 62.4 ± 8.6 years. Table 1 summarizes the clinical characteristics of the training and Validation cohort. All patients were followed up until May 2022, and a total of 39 patients had a death endpoint. The median OS is 7.5 (2–25) months.

Table 1. Demographic and clinicopathological characteristics of patients with PDAC.

Characteristic	Training cohort (N = 32)	Validation cohort (N = 20)
Age (years)	63.3 ± 8.7	60.9 ± 8.4
Sex
Male	17 (53.1%)	13 (65%)
Female	15 (46.9%)	7 (35%)
Tumor location
Head	9 (28.1%)	8 (40%)
Body/tail	23 (71.9%)	12 (60%)
Tumor size (cm)	4.0 ± 1.8	4.6 ± 1.6
CA 19-9 (U/mL)
<1000	15 (46.9%)	9 (45%)
≥1000	17 (53.1%)	11 (55%)
Diabetes
Absent	22 (68.8%)	15 (75%)
Present	10 (31.2%)	5 (25%)
Jaundice
Absent	30 (93.8%)	19 (95%)
Present	2 (6.2%)	1 (5%)
TNM stage
III	17 (53.1%)	5 (25%)
IV	15 (46.9%)	15 (75%)
Vascular invasion
Absent	15 (46.9%)	13 (65%)
Present	18 (56.3%)	7 (35%)
Lymphatic invasion	17 (53.1%)	12 (60%)
Liver metastatic	9 (28.1%)	10 (50%)

Note. Unless otherwise indicated, data are number of patients, with percentage in parentheses.

Radiomic signature construction and evaluation

A total of 1132 radiomics features were extracted from each patient. Finally, through lasso regression analysis, 11 radiomics features (seven from CET1-w arterial phase imaging and four from CET1-w portal venous phase imaging) were selected and used to construct Rad-Score (Figure 2). The evaluation performance of the developed radiomics features was evaluated in two cohorts using the ROC curve. Generally speaking, the diagnostic efficacy of Rad-Score should be affirmed. The AUC values were 0.722 (training cohort) and 0.705 (Validation cohort), respectively.

Figure 2. Radiomics feature selection and coefficients of selected features. (a) Ten-fold cross-validation was applied to select the most suitable feature using the LASSO Cox regression model. (b) Coefficient curves for the 22 parameters.

In addition, according to the optimal cutoff value with the largest Youden index in time-dependent ROC curve analysis, patients from the training and validation cohort were divided into high Rad-Score group and low Rad-Score group. In the training cohort, the prognosis of patients in the low Rad-Score group was significantly better than that in the high Rad-Score group, which was verified in the Validation cohort (Figure 3).

Figure 3. Kaplan–Meier survival analysis was performed according to the optimal cutoff value of Rad-Score of PDAC patient subgroup in training cohort (a) and validation cohort (b).

Radiomics model establishment

Multivariate Cox regression model showed that Rad-Score (HR: 3.224, 95%CI: 1.876–5.541, p < 0.001) and tumor size (HR: 1.515, 95%CI: 1.166–1.968, p = 0.002) were independent influencing factors of OS in the training and Validation cohort (Table 2). Therefore, we constructed a radiomics nomogram model including clinicopathological risk factors and radiomics characteristics (Figure 4(a)). By adding the total score and positioning it on the total score scale, we draw a straight line to estimate the survival probability of each point in time.

Figure 4. (a) Based on the nomogram of radiomics, the 6-, 9-, and 12-month death probability of PDAC cancer was estimated according to the combination of predictors of patients. (b–d) The ROC curves of clinical model, radiomics model, and radiomics nomogram model in training cohort and Validation cohort.

Table 2. Univariate and multivariate analyses of training cohort to identify patient clinical and imaging features with prognostic value for OS.

Variable	Univariate analysis		Multivariate analysis
	HR (95% CI)	p-Value	HR (95% CI)	p-Value
Age	1.001 (0.956–1.050)	0.937
Sex	1.479 (0.662–3.307)	0.339
Tumor location	0.873 (0.359–2.126)	0.765
CA 19-9	1.994 (0.865–4.597)	0.106
Diabetes	2.372 (0.932–6.034)	0.069
Jaundice	0.328 (0.044–2.469)	0.279
TNM stage	1.086 (0.486–2.426)	0.841
Tumor size	3.456 (1.769–6.749)	<0.001	1.515 (1.166–1.968)	0.002
Lymphatic invasion	0.674 (0.296–1.535)	0.347
Liver metastasis	0.858 (0.339–2.175)	0.747
Vascular invasion	0.731 (0.327–1.635)	0.446
Rad-Score	2.696 (1.810–4.017)	<0.001	3.224 (1.876–5.541)	<0.001

Note. Data in parentheses are 95% confidence intervals.

Evaluation performance of the radiomic nomogram

Compared with clinics model or radiomics model, the radiomics nomogram model showed better prognostic discriminant performance in training (AUC: 0.805) and validation cohort (AUC: 0.800) (Figures 4(b–d)). The calibration curve showed that the model prediction of 6-month, 9-month, and 1-year OS has the best consistency with the actual observation (Figure 5). In addition, we used the DCA curve to evaluate whether the nomogram contributes to clinical treatment strategies (Figure 6). DCA analysis showed that the radiomics nomogram model had a larger area under the curve than the clinical model and radiomics model, which showed the application prospect of radiomics nomogram model in clinical decision-making, indicating that the radiomics nomogram would be a more practical clinical tool for predicting the prognosis of patients with PDAC.

Figure 5. Nomogram calibration curve: in the training and Validation cohort, the probability of survival at intervals of 6 months (orange and light green), 9 months (brownish yellow and dark blue), and 1 year (green and purple) was predicted. The 45° gray line represents the reference line, which showed the ideal prediction.

Figure 6. Decision-curve analysis for the different models. The net benefit was shown on the y-axis and the threshold probability was shown on the x-axis. Use of the nomogram model (green line) achieved the highest net benefit compared with the radiomics model (red line), clinics model (yellow line), treat-all strategy (blue line), and the treat-none strategy (purple line).

Discussion

Pancreatic cancer progresses rapidly and has a poor prognosis [17,18], especially for patients who have lost the chance of surgery. Although chemotherapy and radiotherapy have made some progress [19,20], it is still necessary to explore and evaluate treatment options with significant efficacy. In the past ten years, HIFU ablation has accumulated nearly ten thousand treatment experiences in clinical application. Several clinical studies at home and abroad have shown that HIFU combined with chemotherapy or radiotherapy for advanced pancreatic cancer can achieve a higher clinical benefit rate and longer survival than HIFU alone [21–23], and can also play a synergistic role with radiotherapy and chemotherapy [7,24]. However, the high invasiveness of pancreatic cancer determines that its recurrence or progression after the operation is inevitable.

According to previous studies, except for CA19-9 [25,26], a severely restricted biomarker, there is no feasible biomarker to predict PC. Therefore, we try to develop a more powerful preoperative survival model that relies on non-invasive methods, which may help clinicians determine patients’ treatment plans, such as targeted adjuvant or neoadjuvant therapy for some patients with poor prognosis. On the other hand, it might also help identify patients with very poor prognosis who are unlikely to benefit from surgery.

In this study, 1132 radiomics features were extracted from preoperative MRI images of PDAC patients, and 11 prognostic features were identified by non-zero coefficient Lasso regression. Rad-Score was combined with a linear combination of these features, and Kaplan–Meier survival analysis was used to evaluate the prognostic value of Rad score. Our results showed that Rad-Score composed of 11 radiomics features was significantly associated with prognosis in the training and validation cohort (p < 0.01). In addition, in the training and validation cohorts, the area under the ROC curve was higher than 0.70.

The concept of imaging science was first put forward by the Dutch scholar [27] in 2012. The deep meaning of radiomics is to extract a large amount of image information from images (CT, MRI, PET, etc.) in a high-throughput manner to achieve tumor segmentation, feature extraction, and model establishment, and to assist doctors to make the most accurate diagnosis by deeper mining, prediction, and analysis of massive imaging data information. Compared to the gene expression detection based on invasive biopsy, the non-invasive feature of radiomics analysis provides a wide range of applications for clinical patient. Moreover, the biopsy has obvious limitations as the sampling deviation may lead to heterogeneity within the tumor, and radiomics represents a comprehensive evaluation of the whole tumor [28,29]. Furthermore, the preoperative data used to develop risk assessment models are convenient to obtain with almost no additional cost. These advantages and achievements are encouraging and the prospects are promising. Previous studies have shown that, compared with traditional imaging, radiomics help provide important information about tumor heterogeneity, biology, and physiology, which has a broad prospect in evaluating and predicting clinical outcomes and risk stratification. A previous study [30] has shown that several texture features based on T2WI and ADC are useful for the differentiation of prostate cancer. Ma et al. [31] used semi-automatic segmentation of DCE-MRI images to extract radiomics features (morphology, grayscale statistics, and texture features) of 377 women diagnosed as invasive breast cancer. It was found that the quantitative radiomics imaging features of breast tumors extracted from these data were related to the expression of Ki67 in breast cancer. Jain et al. [32] showed that the radiomics features in and around the lung tumor extracted from CT images can predict not only OS but also the response of patients with small cell lung cancer to chemotherapy.

At present, radiomics has been explored in the field of pancreatic tumor research. Benedetti et al. [33] reported that the histopathology of neuroendocrine tumors could be distinguished by the radiomics features derived from CT. Kaissis et al. [16] reported that the radiomics features extracted from preoperative diffusion-weighted imaging sequences can predict OS with high diagnostic accuracy. In our study, the multivariate Cox risk model showed that tumor size was an independent prognostic factor in patients with PDAC which has been confirmed in many previous studies. Comparing with histopathology, Lee et al. [34] concluded that larger tumors were an independent prognostic factor for post-operative decrease in RFS and OS in pancreatic cancer. A preoperative MRI texture analysis study conducted by Choi et al. [13] showed that tumor size and moderately textured entropy were significantly correlated with overall survival in patients with pancreatic ductal adenocarcinoma (PDAC). In addition, it was encouraging that the independent prognostic factor of Rad-Score showed stronger predictive ability in the multivariate COX risk model, which was validated in the validation cohort. Then, we further divided patients into high Rad-Score group and low Rad-Score group. There was significant difference in prognosis between the two groups(training cohort: p < 0.0001, validation cohort: p = 0.0041). Therefore, Rad-Score improves the traditional prognostic ability to rely solely on clinicopathological risk factors, and may provide an effective and promising tool for evaluating the prognosis of PDAC patients.

Recent studies [35–37] showed that the combination of clinicopathological risk factors and Rad-Score can predict patient survival more accurately than relying on Rad-Score alone. Therefore, to improve the accuracy of prognosis estimation, we further developed a radiomics Nomograph combined with clinicopathological risk factors and Rad-Score to improve the accuracy of prognosis estimation. In our current study, higher AUC (training cohort: 0.805; validation cohort: 0.800) and DCA analysis showed that radiomics nomogram could better predict the prognosis of patients at different time points than clinical model or radiomics model. This is consistent with the results of Xie et al. [37], which show that the inclusion of Rad-Score in radiomics nomogram can better predict patient survival than clinical models and TNM staging system. Liang et al. [38] conducted a previous study, the results showed that after adding clinical factors to predict the preoperative histological grading of pancreatic neuroendocrine tumors, the radiomics nomogram showed a significant improvement over the individual radiomics features. This report also confirms our view.

In addition, pancreatic cancer is a highly heterogeneous tumor, and sometimes even the early stage does not necessarily have a good prognosis. This study included patients with locally advanced or advanced stages, and there was no significant difference in the prognosis of these patients. In our study, metastatic disease at presentation was not significantly associated with OS. This may be due to the difference in lesion location between our study and previous studies. In our study, most patients had tumors located in the body and tail of the pancreas. Because tumors in the body and tail of the pancreas often have more invasive tumor biology. Our results are similar to Cheng et al. [39].

Unfortunately, the sample size of this study is small, which is considered to be the most important limitation, because HIFU in the treatment of advanced pancreatic cancer is not yet widespread in China. Although the sample size is relatively small, we found a strong correlation between Rad-Score and OS. In the future, we will continue to accumulate sample size to further explore and correct the prognostic model. Secondly, patients have received different types of chemotherapy, and the heterogeneity of treatment strategies may affect the results. Finally, Rad-Score is calculated by manually drawing regions of interest, which is not only expensive but also may be subjective. Automatic segmentation or semi-segmentation is needed in PDAC radiomics analysis in the future.

Conclusion

To sum up, our results show that Rad-Score is a significant prognostic factor in patients with PDAC. In addition, the radiomics nomogram has great potential in predicting the OS of patients, which helps us better predict the survival of patients after HIFU, thus providing help for clinical practice. Without doubt, for the clinical application of personalized medicine, wider treatment options and better survival are fundamental challenges to this almost invariably fatal disease. With a deeper understanding of imaging biomarkers, hopefully, we would provide more insights into the possible benefits of the new treatment in clinical patients in the future.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.