Author: Pradeep Bhetwal, Yingxuan Chen, Wookjin Choi, Michael Dichmann, Adam Dicker, Rupesh Ghimire, Yevgeniy Vinogradskiy, Maria Werner-Wasik 👨🔬
Affiliation: Thomas Jefferson University 🌍
Purpose: Radiomics has emerged as a powerful tool in medical research. However, the lack of standardized and reproducible pipelines limits its clinical adoption. This study developed a robust and scalable pipeline for streamlined radiomics analysis in a clinical setting, enhancing the integration of radiomic features into patients with lung cancer receiving radiation therapy.
Methods: The proposed pipeline automates the radiomics workflow, including data import, preprocessing, feature extraction, post-processing, and prediction model building and inference. Planning datasets for radiotherapy DICOM files are retrieved via the MIM Radiation Oncology PACS, aligning planning CT scans with their corresponding RTPLAN, RTDOSE, and RTSTRUCT files. The pipeline converts CT images, target structures, organ-at-risk structures, and dose maps for radiomic feature extraction. We used SimpleITK for medical images and their metadata process, PyRadiomics for radiomic feature extraction, and Prefect for task orchestration for smooth error handling, scheduling, and failure management. The pipeline was applied to data from 207 lung cancer patients treated from 2019 to 2022, including 340 gross tumor volumes (GTVs) from 226 CT scans. Validation was done by the TCIA NSCLC-RADIOMICS dataset including manual GTV delineations from 422 non-small cell lung cancer (NSCLC) patients.
Results: The pipeline successfully processed both datasets, the institutional dataset (275 tasks, 1,479 targets) in 3,872 seconds, averaging 14.08 seconds per task and 2.62 seconds per target, with 27 failures. The NSCLC-RADIOMICS dataset (422 tasks, 728 GTVs) took 1108 seconds, averaging 2.63 seconds per task and 1.52 seconds per target, with one failure.
Conclusion: This study developed a scalable and robust radiomics pipeline validated across clinical and public datasets. The pipeline enhances the potential for precision medicine in lung cancer care by seamlessly integrating radiomics analysis into clinical workflows. Future work will incorporate clinical data to explore the prognostic and predictive value of radiomic features for lung cancer treatment outcomes.