Purpose/Objective(s)Differentiation between radiation-induced lung fibrosis (RILF) and tumor local recurrence (LR) remains challenging in SBRT treated NSCLC patients. We aimed to integrate spatially and time-resolved 4D radiomics with dosiomics biomarkers to develop a novel robust multi-omics classifier of LR vs RILF.Materials/Methods210 NSCLC patients (101/48% T1-2N0M0, 109/52% T1-3N1/xM0/x) treated with SBRT (median dose 60 Gy/8 fractions) between 2009 and 2019 were identified. Image-based quantitative features were extracted from the SBRT planning and follow-up (FU) CTs at the time point of diagnosis of LR/RILF or a matched time point for CTs without LR/RILF (radiomics) and from dose distributions (dosiomics). The region of interest was the planning target volume plus a 10mm margin (PTV+10mm). Time-dependent alterations of the radiomics features in FU CTs were integrated (delta-radiomics). To identify relevant features associated with RILF or LR, resampling (iterations = 1000) of feature selection methods were applied. An ensemble classifier comprising random forests, neural networks, and logistic regression was then trained. The area under the ROC curve (AUC) with a 70%/30% training/testing split and a 5-fold cross-validated AUC on the entire cohort were calculated for performance assessment at CT FU.ResultsOut of the 210, 36 patients (17%) were reported with LR and 44 (21%) with RILF. The median FU time was 19 months (range, 5-87). 20 LR and 23 RILF FU CTs, both with median time after SBRT of 15 months, were identified. 53 patients with matching clinical characteristics without LR/RILF (none), at matched FU time points (median time after SBRT: 17 months), were additionally selected, resulting in a cohort of 96 patients (20/21% LR, 23/24% RILF, 53/55% none). Discrimination of RILF versus LC using 4D radiomics features was achieved with a testing and 5-fold AUC of 0.82 [95%CI 0.79 0.86] and 0.85 [0.83 0.88]. The addition of a dosiomics feature improved performance to 0.85 [0.82 0.87] and 0.88 [0.85 0.91]. The PTV+10mm-based classifier includes 4 textural features (2 delta-radiomics, 1 radiomics from the FU CT and 1 dosiomics). A non-significant correlation was found between the significant features and tumor volume before RT (Spearman's r <0.1, p<0.05).ConclusionOur data indicate that integrative omics by combining radiotherapy volume and dose constraints with spatially and time-resolved radiomics may provide a novel mean for better discrimination of tumor recurrence vs RILF after SBRT.

Personalized medicine in Radiotherapy (RT) aims to increase tumor control probability and decrease normal tissue toxicity. Recently, Radiomics is widely used to infer tumor/tissue characteristics and link outcome to applied dose and underlying biology. The building of robust models requires high amounts of data, which are often only available from multiple centers/sites. Missing standardization of data (inconsistent naming schemes, differences in acquisition parameters) makes manual curation necessary, which can be extremely time-consuming. Here, we propose a fully automated general framework that can sort all relevant RT data with minimal human intervention - pyCuRT. Our method builds upon NyPipe, a Python package used to create complex analysis workflows. Any DICOM directory containing different radiological data e.g., batch exported from PACS systems without specific structure requirements, can be used as input for pyCuRT. It checks the integrity of files and sorts them based on DICOM attributes. For RT data, using information from the RT Plan, pyCuRT links together the DICOM-RT objects, i.e., the RT planning CT, Structure Set (SS), and Dose Distribution (DD). Furthermore, the structure within the SS showing the highest overlap with the DD can be automatically identified, allowing the extraction of inconsistently named structures. The final output has a subject/session/scan structure. Parallelization is implemented to speed up computation on multi-core machines. Successful curation has been achieved in retrospectively collected data cohorts of 3 different cancer entities (brain n = 621, rectal n = 127, and pancreatic n = 13 for a total of ∼50000 scans), from several various institutions across Germany. In the rectal cohort, e.g., from more than 2000 present images, 774 CT scans with 224 unique series descriptions were identified. From them, pyCuRT extracted the RT planning CTs, and linked them to the corresponding RT-DD and RT-SS automatically. Furthermore, all structures with any combination of GTV or PTV in the name were correctly extracted from the SSs and saved as NIFTI (the 35 RT-SS not containing those structures were identified correctly). The total processing time was around 2 hours and a half, which corresponds to less than 3 seconds per scan. We propose a new, fully automated method to curate radiotherapy data coming from different institutions and vendors. Its robustness and usability were demonstrated in three cohorts of tumor entities treated with RT. Future work will expand the workflow to automatically classify MR sequences.