AAI_2025_Capstone_Chronicles_Combined

6

further illustrate how differently clinicians delineate gross tumor volume (GTV), reflecting well-documented challenges in achieving reproducible target definition (Wee et al. 2019). Extensive clinical evidence confirms that both intra- and inter-observer variability are significant sources of uncertainty in lung cancer contouring, affecting staging, treatment planning, and response assessment (Van De Steene et al. 2002; Das et al. 2021). A systematic review reinforced these concerns, highlighting delineation variability as one of radiation oncology’s most persistent and consequential sources of error (Vinod et al. 2016). Effects of CT Acquisition Protocols In addition to labeling variability, differences in CT acquisition contribute substantially to volumetric inconsistencies. Radiomics studies have shown that image-derived features are sensitive to reconstruction kernels, slice thickness, and scanner-specific parameters (Aerts et al. 2019). Deep learning–based conversion of reconstruction kernels has been shown to improve the reproducibility of radiomics features for pulmonary nodules and masses, underscoring that acquisition variability can meaningfully influence quantitative imaging analyses (Choe et al. 2019). Complementary studies evaluating volumetry have observed that measurement precision varies with lesion morphology and other nodule characteristics, and that diameter-based approximations can differ substantially from voxel-based estimates, particularly for irregular tumors (Hwang et al. 2017; Yoon et al. 2022). Together, these findings highlight the need for segmentation and volumetry methods that are robust to

Figure 2.1​ Per-slice area by position along the normal, with an example of a spiculated mass with non-tissue cavity Key Takeaways The NSCLC-Radiomics dataset provides the essential components required for accurate, reproducible 3D tumor volumetry: expert-defined GTV contours, fully specified DICOM geometry for reconstructing masks in patient space, and substantial variability in tumor size and morphology to rigorously test model robustness. These properties make it an appropriate and reliable benchmark for evaluating the performance of 3D segmentation and tumor volume inference models. 3 Literature Review Segmentation Variability and Ground-Truth Limitations Accurate quantification of lung tumor volume from computed tomography (CT) remains a longstanding challenge in thoracic oncology due to substantial interobserver variability, heterogeneity in imaging acquisition, and the inherently irregular geometry of lung tumors. Early radiomics studies using datasets such as NSCLC-Radiomics and LUNG1 demonstrated the promise of quantitative imaging biomarkers for prognosis but also emphasized the dependence of these methods on consistent and well-defined segmentation inputs (Aerts et al. 2019; Braghetto et al. 2022). Multi-reader datasets containing multiple expert contours

403