The purpose of this study was to explore the predictive value of radiomics features from dual-time images on the prognosis of LAPC patients treated with SBRT. We found that the Rad-score from dual time 18F-FDG PET/CT images can be used to predict the prognosis of LAPC patients after SBRT, and can achieve the prognostic stratification with OS of patients. This result has clinical significance for promoting the precise treatment of LAPC.
The radiomics has attracted a lot of attention in its ability to non-invasively analyze tumor heterogeneity; and provides a viable tool for patient prognosis. In this study, we conducted a two-stage experimental setup to explore the best model for a prognosis for patients with pancreatic cancer. First, we use 594 radiomics features extracted from early imaging or delayed imaging to develop a single-time prognostic model, and then we combine the radiomics features extracted from dual time PET/CT images to develop a dual-time prognostic model. In these two stages, the six radiomics features with the highest frequency were used to calculate the Rad-score. Univariate analysis showed that the Rad-score of both early and delayed images was significantly correlated with the OS of LAPC patients (p < 0.01, HR: 2.60 and 2.70). This result indicated that both early and delayed PET/CT images contain patient prognostic information[12,22]. The Rad-score from dual time images is also significantly correlated with OS of patients (p < 0.01, HR: 2.35). The C-index of the model based on dual time images was significantly higher than that of the model based on early and delayed images (p < 0.05, Wilcoxon rank-sum test). At the same time, it must be pointed out that there are differences in the division of high-risk and low-risk groups according to the Rad-score of single time images, while an accurate grouping can be obtained by analyzing the Rad-score of dual time images. Two representative patients were shown in Fig. 3. These results showed that dual-time PET/CT images can provide more prognostic information and offset the limitations of single-time PET/CT images. This may be related to the kinetics of tumor uptake. As the uptake time increases, the uptake of 18F-FDG by malignant tumor tissues will increase significantly[31–32].
In this study, the Rad-score from dual time images includes three types of features: shape feature, first-order gray statistics feature, and texture feature. Different types of texture features can reflect the inherent heterogeneity of tumors from different angles[33–34]. In addition, the optimal features include both original texture features and texture features based on wavelet transform (wavelet coefficients: LLH + LHL + HLL, HHH (HHH = highpass filter + highpass filter + highpass filter)), which reflect the heterogeneity of tumors on different spatial scales. On the one hand, it must be pointed out that when these six radiomics features were used alone, none of the features can predict patient survival better than the Rad-score, indicating the complementarity of information between different features. On the other hand, these characteristics are also significantly related to the OS of patients. The solidity reflects the complexity of the homogenous region of the tumor. The smaller the value, the higher the complexity of the tumor and the worse the prognosis of the patients. The gray level difference statistics (GLDS) calculates the contrast of the image and reflects the roughness of the texture. The Contrast  (PET, GLDS, HHH) shows that the metabolic changes in the lesions in PET images have a strong resolution, the functional metabolic changes in the lesions are larger and the texture is coarser in patients with poor prognosis. The Busyness (Neighborhood Gray-Tone Difference Matrix (NGTDM), PET) measures the change from pixel to the adjacent pixel. A high value of business indicates that the intensity between a pixel and its neighborhood changes rapidly, indicating that the more complex the tumor is in patients with poor prognosis. The Correlation and The Energy from Gray Level Size Zone Matrix (GLSZM) in CT images quantify the degree of non-uniformity of the gray level in images. The patients with a better prognosis showed better texture consistency.
For LAPC patients, the median survival was 4–16 months. The clinicopathological parameters, including T-stage, chemotherapy, and dose, were found to be strong predictors of prognosis in LAPC patients receiving SBRT in univariate Cox analysis. However, the above indicators did not show prognostic value for patients in the multivariate Cox analysis study. For the conventional PET features, the TLG from delayed PET/CT images is significantly associated with a poor prognosis. This reaffirms the fact that the metabolic tumor volume combined with the tumor range is a better predictor of patient survival. However, in this study, the TLG of the early images does not correlate with the OS. One possible reason is that compared with the early images, delayed images may better reflect the uptake of 18F-FDG by malignant tumors. This result is different from our previous research. One possible reason is that the tumor contour is outlined in different ways, and another possible reason is the difference in the number of patients.
There are some limitations to this study. First, the samples in this study were from a single-center, the sample size available for analysis was small, and the potential of selection bias cannot be ruled out, which limits the accuracy and reliability of the results. Therefore, we hope that the results of this study can be repeated using larger data sets and multiple centers in the future. Second, the ROI/VOI was drawn manually, which is very time-consuming and inconvenient, and the predicted performance may be sensitive to the ROI/VOI depicting pancreatic lesions. In future research, automatic segmentation or semi-automatic segmentation could be achieved through the application of deep learning. Finally, PET/CT images acquisition was carried out in the presence of respiratory movement, which may distort the real metabolic activity and affect the accurate segmentation of tumors. In future research, the respiratory gating technique may help to improve the quantitative accuracy of PET imaging.