ESTRO 2025 - Abstract Book

S3043

Physics - Image acquisition and processing

ESTRO 2025

Purpose/Objective: To develop a Python library based on ITK and SimpleITK [1–3] for reading/writing medical images in different formats, including DICOM and NIfTI. The library was then applied to compare Radiomic Features (RFs) generated by the MATLAB version of the highly IBSI-consistent [4] SPAARC software (https://spaarc-radiomics.io/ [5, 6], https://theibsi.github.io/) (which requires DICOM inputs) against Pyradiomics [7] (compatible with ITK formats, including NIfTI). Material/Methods: Existing libraries were preliminarily considered, but none met our optimal criteria: in particular the conversion from DICOM to NIfTI was incorrect. The rt-utils library [8] was considered for NIfTI-to-DICOM conversion, but produced structural discrepancies when compared to original DICOM data. A Python 3.10 library was created to maximize compatibility with ITK, emphasizing maintainability and future usability. The documentation was made publicly available at https://github.com/pymaitre (MIP library). The library was validated against diverse data structures, including those with complex topologies (e.g., multiple regions or internal cavities). For the comparison of RFs, 158 CT scans from three different scanners of pancreatic adenocarcinoma patients were utilized. A set of 102 RFs was selected, excluding features exclusive to either SPAARC or Pyradiomics [8]. The relative differences between SPAARC and Pyradiomics were calculated by normalizing to the average feature value. Intraclass Correlation Coefficients (ICCs) [9] were computed to assess whether predictive models remain robust independently from the software used for RF extraction. A threshold ICC of 0.99 was used to classify features as ”robust”. Results: As shown in Figure 1, the generated NIfTI masks are well within the original DICOM structure, and the saved DICOM structure perfectly matches the original one.