CT-based imaging metrics for identification of radiation-induced lung damage

PurposeThis study investigates the feasibility of radiomics for identifying textural changes of radiation-induced lung damage (RILD) after chemoradiotherapy. MethodsThe severity of RILD on each CT scan was graded on a scale from 0 (similar to the baseline CT scan) to 5 (lung fibrosis). The delineation of abnormal areas inside the lung on CT images was performed semi-automatically using a median filter. We extracted a total of 138 quantitative image features from this delineated region of interest and ran a random forest algorithm as a classifier for identifying the severity of RILD. After training and testing the model, we validated the model using a separate dataset. ResultsThe classification accuracies for identifying grade 0 from grades 1 [~] 5 were 70% for the test dataset and 85% for the validation dataset; for identifying grade 1 from grades 2[~]5, 90% for the test dataset and 95% for the validation dataset; and for identifying grade 5 from grades 2[~]4, 80% for the test dataset and 85% for the validation dataset. ConclusionsOur preliminary study shows that the classification accuracy was robust, the model was most useful for distinguishing grade 1 from other grades, and the results demonstrated the feasibility of radiomics for identifying the severity of lung damage after chemoradiotherapy. This approach could be a potential tool for helping diagnostic radiologists identify RILD and its severity on CT images.