Purpose: To investigate the feasibility of predicting the occurrence of radiation pneumonia (RP) in patients after radiation therapy (RT) of breast cancer using dosiomics index.
Methods: A pilot retrospective study of 42 breast cancer patients was performed under an institutional IRB protocol. 11 of 42 patients were observed with RP complications based on CT scans 3-5 months post RT. We hypothesize that the dosiomics features can better characterize the occurrence of RP than the traditional volume-based dosimetric metrics (such as V5, V20, and mean dose). To test, we first evaluated the potential dosimetric deviations due to inter-fractional mispositioning and intra-fractional patient respiratory motion using daily cone-beam CT (CBCT). The actual dose distribution within the lung was re-calculated by incorporating the CBCT images taken before each fractional treatment. 91 dosiomics features were extracted from both the whole lung dose and the high dose region (>20 Gy), respectively. T-test was performed to find the optimal features as potential indexes for RP. The comparison of p-value between the dosiomics features and the traditional dosimetric indexes was also conducted to evaluate the overall performance.
Results: Dosiomics feature GLCOM-based Sum Variance extracted from whole lung dose presents the significant statistical difference between 11 RP patients and rest 31 patients (p-value = 0.0487), and feature GLCOM-based Cluster Shade extracted from high dose region gives lower p-value (0.0316). While the traditional dose constraints are insufficient to characterize the dose distribution differences between two groups of patients (p-value for mean dose = 0.5447, p-value for V5 = 0.7926, and p-value for V20 = 0.4058).
Conclusion: This pilot work verified the feasibility of evaluating the occurrence of RP in patients after breast cancer RT with dosiomics features. The dosiomics-based indexes, feature GLCOM-based Sum Variance and Cluster Shade, showed the potential to replace the traditional dosimetric indexes in predicting RP.