Preoperative prediction of pathological grade in pancreatic ductal adenocarcinoma based on ¹⁸F-FDG PET/CT radiomics

Pancreatic ductal adenocarcinoma (PDAC) is the fourth leading cause of cancer-related death worldwide, which accounts for about 85% of all pancreatic tumors [1, 2]. Surgical resection is the only curative treatment for PDAC, and the addition of chemotherapy in the adjuvant setting has been shown to improve survival rates [3]. Despite the advances in diagnostic and therapeutic modalities, the prognosis of PDAC remains poor, with a reported 5-year survival rate of 7.2% [4]. Therefore, it is crucial to identify the prognostic factors that may predict the prognosis of patients with PDAC, which can help clinical physicians select patients for the available treatment options.

Pathological subtype is considered to be a crucial factor in PDAC prognosis, and the presence of poor differentiation portends a poor prognosis [5‐7]. Intraoperative frozen sections are used for the patients undergoing pancreatectomy; however, this method is time-consuming and expensive relatively. For the patients with non-resectable tumor, endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) is a common approach to get the specimen of the tumor for pathological examination before making a treatment plan [8]. Nevertheless, this technique is highly invasive, which has the inherent risk of interventional complications, and the achievable samples are too limited to give a reliable histological grading [9]. The accuracy of EUS-FNA to determine the grade of tumor is still challenging in clinical practice [10, 11]. There is a need for a reliable technique to evaluate the differentiation of tumor before treatment.

Noninvasive imaging tools are playing a vital role in the preoperative evaluation of PDAC. Triphasic computed tomography (CT) is the best initial diagnostic method for PDAC, and magnetic resonance imaging (MRI) can be used in patients who cannot tolerate the intravenous contrast for CT [7]. Compared to traditional imaging methods, ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography/computed tomography (PET/CT) can combine functional information and anatomic information. Over the past 20 years, studies have shown the potential of ¹⁸F-FDG PET/CT in the diagnosis, staging, and recurrence of PDAC [12]. ¹⁸F-FDG PET/CT parameters, including standardized uptake value (SUV), metabolic tumor volume (MTV), and total lesion glycolysis (TLG), have been reported as prognostic factors for pancreatic cancer [13‐15]. However, established ¹⁸F-FDG PET/CT parameters are difficult to assess the histopathological differentiation of tumor [16]. The ability of PET/CT to noninvasively evaluate the differentiation of tumor has not been fully explored.

Radiomics, which was initially proposed by Lambin et al. in 2012 and defined as high-throughput extraction of large amounts of features from radiographic images, has the ability to capture intratumoral heterogeneity noninvasively [17, 18]. Radiomics has the potential to capture information beyond what is visible to the human eyes in an objective and repeatable way [19]. For the last decade, an increasing number of studies have been conducted to assess the efficacy of radiomics for the characterization of tumor heterogeneity [20]. Texture analysis can reflect the intratumoral heterogeneity to help assess tumor aggressiveness, and lower tumor heterogeneity is likely related to lower histological differentiation [21, 22]. Previous studies suggested that features derived from CT, MRI, and PET/CT favor the tumor classification or survival prediction [10, 23‐26].

To our knowledge, there have been a few published reports regarding the value of PET/CT-based texture analysis for evaluating the histopathological differentiation of pancreatic tumor. Therefore, the purpose of our study was to develop a machine-learning model based on radiomic features extracted from PET/CT images for predicting the pathologic differentiation of PDAC preoperatively.

An overview of the study workflow is illustrated in Fig. 1, and the radiomics process is divided into five steps: region of interest (ROI) segmentation, radiomics feature extraction, feature selection, radiomics-based model construction, and model evaluation.

The purpose of this retrospective study is to investigate the value of PET/CT radiomics for predicting the risk assessment of PDAC. In this work, we present a model based on 12 radiomics features found to be effective for the classification, which could stratify PDAC patients into grade 1 and grade 2/3 groups with AUC of 0.994 in the training set and 0.921 in the validation set. The results suggested that PET/CT radiomics with machine learning may contribute to the preoperative risk assessment of PDAC patients.

Differentiation grade is an important biologic feature of malignant tumors. Although pathological examination is a necessary exam to determine the histological classification of the tumor, radiomics might have an increasing clinical role, as it provides an atraumatic in vivo and simple method to characterize the lesions. Some scientists have explored the value of PET-radiomics in tumor classification. Hyun et al. successfully identified the histological subtypes of lung cancer using a machine-learning algorithm with PET-based radiomic features [31]. Also, Han et al. showed that machine learning/deep learning algorithms could help radiologists to differentiate the histological subtypes of non-small cell lung cancer via PET/CT images [32]. Zhang et al. confirmed that the quantified radiomics method could aid the noninvasive differentiation of autoimmune pancreatitis and PDAC in ¹⁸F-FDG PET/CT images [26]. Similar to the above-mentioned studies, current one also supports the conclusion that PET/CT texture analysis is a promising method for predicting the differentiation or subtype of the tumor.

Several studies have evaluated the potential value of radiomics for predicting pathology of pancreatic tumors. Attiyeh et al. present a multivariate model of quantitative imaging features and clinical variables that predicts low and high risk in intraductal papillary mucinous neoplasms [33]. Kaissis et al. developed a machine learning algorithm from diffusion-weighted imaging-derived radiomic features, which has the potential in PDAC preoperative subtyping [24]. Furthermore, Liang et al. developed a nomogram model combing radiomics features and clinical characteristics, and it showed the best performance in differentiating grade 1 and grade 2/3 pancreatic neuroendocrine tumors [10]. Recently, Qiu et al. found that there were significant differences between low-grade PDAC (well-differentiated and moderately differentiated) and high-grade PDAC (poorly differentiated) regarding 18 CT texture features [34]. These studies are mostly based on CT and MRI modality. We collected PET/CT images as the image modality in this study. CT or MRI reflects mainly the anatomical structure of tumors, while PET/CT can be used to explore intratumoral heterogeneity in both anatomical and functional dimensions [35]. In our study, the ¹⁸F-FDG PET/CT-based texture analysis demonstrated excellent performance in identifying the pathological grade of PDAC.

The current study showed that CT radiomics features are better in predicting the pathologic grade of PDAC than PET features. This finding is as expected since PET images has lower resolution than CT images, and previous studies have indicated that the performance of texture analysis is better in the analysis of CT images than PET images [36]. Kirienko et al. found that the added value of PET appears limited when comparing the AUC of radiomic signature among CT, PET, and PET + CT in the prediction of disease-free survival of non-small cell lung cancer [37]. However, there are very limited literatures on the role of radiomics of both PET and CT images in pancreatic tumor. Our study shows that the model using both the PET and CT features has an advantage compared to using either PET or CT features alone in predicting the pathologic grade of PDAC. Also it is worth noting that texture analysis of CT images was performed using the CT component of the PET/CT instead of diagnostic CT images, so evaluating the predictive performance of PET/CT and diagnostic CT is an interesting and valuable research direction. Besides, we attempted to divide patients into training and validation set in another chronological order in the ratio of 2 to 1, but the distribution between two sets was not even (Additional file 1: Fig. 1 and Table 1), with the lymphatic metastasis demonstrated significant difference between the training and validation set (p < 0.001). Thus, we did not adopt this grouping method. In the studies of radiomics, validation is an important part in assessing the performance model. It is necessary to fully consider the influence of training data and validation data. Valid models should exhibit statistical consistency between the training and validation sets [21]. Furthermore, externally validated model has more credibility than internally validated model, because data obtained with the former approach are considered more independent [38].

It is interesting to note that in this study, features extracted from wavelet filtered images played a significant role in the prediction model. Wavelet transform can decompose images into low- and/or high-frequency components at different scales, and the intensity and textural features exacted from wavelet decompositions of the original image can represent different frequency ranges within the tumor volume. Previous studies have shown that wavelet transform-based radiomic features can serve as independent predictive or prognostic factors [39, 40], and are correlated with histological classification results and/or genes mutation status [41, 42]. Our study also suggests the value of wavelet features in identifying the pathological grade of PDAC.

As a retrospective study, a limitation of our study is that the number of enrolled patients was relatively small. Because of the sample number of poorly differentiated cases was small, grade 2/3 PDAC group was not further separated into moderately and poorly differentiated cases. Also, the present study depends on the pathological diagnosis; thus, our cohort was restricted to resected patients, and selection bias may exist implicitly. In the future, we will expand our patient cohort and validate the ability of this prediction model to separate well, moderately, and poorly differentiated patients in a prospective study.

In conclusion, this model based on PET/CT radiomics and machine learning can predict the histological differentiation of PDAC effectively and preoperatively. Before it can be used in clinical practice, prospective studies on larger patient cohorts are required to validate its utility further. As the development and standardization of PET/CT radiomics, it could be a promising preoperative, noninvasive, precise, and simple approach that can assist physicians in evaluating the risk of patients with PDAC individually, and thus achieving a personalized treatment and a better clinical outcome.