RP is a very important adverse event in radiotherapy for NSCLC, and if it can be predicted, it will be useful information for deciding the treatment policy, such as prescription dose and follow-up interval.
In the past, grade 3 RP was applied to predict RP. Dang et al. reported that grade 2 and grade 3 RP have different predictors . Therefore, predictors of grade 2 RP require an original approach. Since immunotherapy has revolutionized the treatment of lung cancer , it is very important to predict grade 2 RP because it prevents the continuation of immunotherapy. In the current study, we focused on finding a new method to predict grade 2 RP. A previous report showed that background factors such as sex, smoking status, tumor location, age, and pulmonary comorbidity have been identified as potential risks [8, 9]. Dosimetric factors, such as the mean lung dose or V20, and other several DVH parameters have also been reported as the best correlated predictors of RP. It is necessary to develop predictors that take both these factors into consideration.
Currently, radiomics approaches have been used to improve diagnostic quality or to predict treatment outcomes by using medical images that have only been used for radiation diagnosis, treatment planning, and follow-up after treatment. Krafft et al. combined radiomics features extracted from whole-lung images with clinical and dosimetric features and significantly improved the RP model for grade 3 RP . The cross-validated AUC of their model was 0.68. The current study compared the prediction model using the whole-lung radiomics analysis that was performed by Krafft et al. and multi-region radiomics analysis, which was proposed as a new method in the current study. We were able to reproduce the same degree of accuracy by using a method like that of Krafft et al., as though we predicted grade 2 RP rather than grade 3. Furthermore, we could improve the quality by using multi-region radiomics, which had superior accuracy, sensitivity, specificity, and AUC compared to whole-lung radiomics analysis.
In the multi-region radiomics analysis, the radiomics features of local intensity roughness and variation were selected from the CT image. Previous study showed that the increasing in radiologic density within the irradiated lung was to be a predictor for the RP. Thus, the radiation pneumonitis can be occurred in local intensity roughness and variation can cause increasing the density on the CT image, which is important factors in predicting RP. Moreover, the shape features were extracted from dosimetric-based segmentation, which reflect dose-volume histogram metrics. Thus, dosimetric parameters are essential for the prediction of the RP grade. For radiomics features in the dosimetric-based segmentation, the regions of the normal lung that received 60 Gy were selected as an important predictor in addition to the region of the normal lung that received 20 and 30 Gy. Although there are many reports that emphasize the importance of the median , the correlation between a high dose and RP has not been fully analyzed. We previously reported the importance of reducing high doses by analyzing NSCLC patients who received 3D-CRT or intensity-modulated RT (IMRT) . The current study supports the importance of the higher dose as well as the median dose. We should reduce not only low and middle doses but also higher doses to prevent RP.
The problem with the previous radiomics analysis is that it is calculated from one value in the segmented region. RP mostly occurs in the locally irradiated lung. Multi-region analysis can extract radiomics features from the locally irradiated region in addition to the whole lung. Hao et al. proposed a predictive model for distant failure by shell analysis. Shell analysis extracts features in the boundaries of the tumor, which allows us to detect its associations with metastases within the microenvironment . The current study also demonstrated that the radiomics features extracted in the multi-locally segmented region could be an important predictor of the classification of above and under RP grade 2 .
For feature selection with LASSO regression, all radiomics features extracted by texture analysis were used with the wavelet filter. Nie et al. reported that radiomics features with high-order filters and wavelets were significant predictors for differentiating focal nodular hyperplasia from hepatocellular carcinoma in the liver . In the current study, all radiomics features with wavelet filters that were selected using LASSO regression were important predictors for the prediction of the RP grade. The imaging filter can denoise, smoothen, or enhance edges, and extract or eliminate a constant frequency. This leads to elimination of redundant and effective factors for prediction.
There were some limitations to the current study. Dosiomics and clinical factors such as smoking history, age, and chemotherapy could be integrated into the model to improve its prediction ability and robustness. Liang et al. proposed a prediction model for RP grade using dosiomics analysis from dose distribution for radiotherapy response prediction . Although the multi-region radiomics analysis has improved the prediction ability compared to dosiomics analysis, dosiomics analysis can extract spatial features such as local intensity variation of the dose distribution and the ratio of the low-dose region. Future studies will be performed with a combination model of multi-region radiomics and dosiomics. Another limitation was that the current study used a dataset from a single institution of patients who underwent 3D-CRT. In the future, we will conduct a larger multicenter study using both 3D-CRT and IMRT data to construct a highly versatile predictive model. Ambiguity in the definition of grade 2 pneumonitis may be another limitation as in the past report. In previous reports, the incidence of grade 2 pneumonitis varied considerably, and one of the causes may be that the definition was different in each paper. In this study, the definition of grade 2 RP was based on CTCAE ver.5.0. We tried to minimize subjective judgment by defining it as RP that need some kind of treatment. Still, we successfully combined both individual patient background factors and dosimetric factors by analyzing RP with our new Radiomics methods. Based on the prediction method developed in this study, it may be possible to reexamine the treatment plan by analyzing the images of the treatment plan and predicting the risk of Grade 2 pneumonitis before the start of treatment.