AAI_2025_Capstone_Chronicles_Combined

8

barriers to robust performance (Hiraman et al. 2024). A smaller but growing body of work focuses directly on volumetry and 3D reconstruction. One CT-based study proposed a 3D reconstruction framework incorporating adversarial and transductive learning to improve tumor shape consistency across domains, illustrating the potential benefits of modeling volumetric topology more explicitly (He et al. 2025). Reviews of AI-driven volumetry tools in nuclear medicine have similarly noted that, while many systems report strong correlations with expert-generated volumes, evaluation practices vary widely, and standardized validation - particularly for 3D geometric accuracy - is still limited (Wendler et al. 2023). Importantly, many segmentation-focused studies treat segmentation as the endpoint, even though tumor volume - not segmentation itself - is often the clinically actionable quantity. (Rutkowski, 2014) Parallel developments in other modalities reinforce the importance of volumetric accuracy. Several studies have demonstrated the feasibility of automated metabolic tumor volume (MTV) estimation in PET imaging for lymphoma and other cancers, with deep learning–based methods producing MTV and total MTV (TMTV) measurements that correlate closely with expert annotations (Kuker et al. 2022; Jemaa et al. 2019). In CT imaging of abdominal organs, related work has shown that deep learning–based volumetry is feasible for tasks such as kidney and renal tumor segmentation, highlighting the broader applicability of volumetric reconstruction frameworks (Hsiao et al. 2022). Research in liver metastasis segmentation has shown that

tumor volume provides clinically relevant information, further demonstrating that volumetric biomarkers can enhance cancer evaluation across disease sites (Wesdorp et al. 2021). Although these studies differ in modality and clinical context, they collectively support the notion that accurate, automated volumetry is a viable and increasingly important capability in medical imaging. Commercial radiation therapy systems - including tools integrated into RayStation, Varian Eclipse, Elekta planning software, and standalone products such as Limbus Contour - have adopted automated segmentation to improve workflow efficiency. A systematic review of auto-contouring approaches highlights substantial heterogeneity in evaluation methodology and reporting, indicating that many tools have limited publicly documented performance across diverse imaging scenarios (Sherer et al. 2020). While such systems offer practical benefits, their volumetric accuracy under real-world clinical variability remains only partially characterized in the peer-reviewed literature. Taken together, the current body of research defines a clear landscape of methodologies addressing lung tumor segmentation and volumetry. Manual and semi-automated methods remain challenged by observer variability and sensitivity to imaging conditions; deep learning segmentation models offer improved accuracy but face generalizability limitations and inconsistent volumetric reporting; and volumetry-specific research underscores the importance of rigorous geometric reconstruction grounded in DICOM metadata. Within this context, 3D CNN frameworks such as nnU-Net are particularly

405