Radiomics for early detection of tumour recurrence in patients treated with lung stereotactic ablative radiotherapy (SABR)

For early-stage and medically inoperable non-small-cell lung carcinoma (NSCLC) patients, stereotactic ablative radiotherapy (SABR) is a recommended treatment alternative to surgery, with comparable local control rates. Due to the high radiation dose delivered using SABR, patients often experience a radiation-induced lung injury called fibrosis, which can conceal a developing local recurrence on surveillance imaging scans. Identifying recurrences early and accurately is especially crucial if the patient may need an additional therapy to improve their clinical outcome. In this study, radiomics, an emerging medical image analysis tool, is investigated to determine if it can be used to accurately differentiate a local recurrence from fibrosis. This radiomics study is divided into two parts: first the radiomic features proposed in a previous research study were reevaluated in our patient dataset. The preliminary results showed that the performance of a model based on those radiomic features was not reproducible for the follow-up images from our patient set. Second, a new set of features were proposed and validated using three independent validation sets at different time points of follow-up imaging. Post-treatment computed tomography (CT) scans (n = 168) of 79 patients treated with SABR were included. The training set consisted of 24 months median follow up with 16 recurrences and 31 non recurrences (47 patients in total). Independent validation sets were separated into three cohorts by follow-up CT scans at 24 months (cohort 1), 9 – 12 months (cohort 2), and 5 – 9 months (cohort 3). The top five radiomic features selected from the training set were Dependence Non-Uniformity Normalized, Small Dependence Low Gray Level Emphasis, Run Length Non-Uniformity Normalized, Dependence Non-Uniformity and Small Dependence Emphasis. For validation cohort 1 the true positive rate (TPR) was 70% and true negative rate (TNR) was 100% (AUC = 0.85), for cohort 2 the TPR was 55.6% and TNR was 94.4% (AUC = 0.75), for cohort 3 the TPR was 70% and TNR was 90% (AUC = 0.80). Consequently, we’ve demonstrated that radiomic analysis could achieve high TNR in the detection of a non-recurrence and satisfactory TPR in detecting a recurrence. External validation and further study are required to determine the generalizability of this radiomic modeling.Science, Irving K. Barber Faculty of (Okanagan)Computer Science, Mathematics, Physics and Statistics, Department of (Okanagan)Graduat