AAI_2025_Capstone_Chronicles_Combined

13

single-fold, partially trained checkpoint rather than as the endpoint of the full nnU-Net training protocol. Within those limits, the nnU-Net v2 model demonstrated clear learning behavior: training and validation losses declined, and validation pseudo Dice improved over time. However, when evaluated on a held-out test set, segmentation performance remained low for many cases, and volume estimates often deviated substantially from the clinical reference volumes, typically in the direction of overestimation. Under the available compute and time, the model did not yet achieve the level of generalization that would be required for clinical or research deployment as an automatic volumetry tool. The more durable contribution of this work lies in the pipeline itself. The project established a DICOM-native, geometrically consistent process that links CT image series, RTSTRUCT contour data, voxel-aligned masks, and volume calculations into a single, auditable workflow. This framework makes it straightforward to substitute alternative models, adjust training regimes, and generate comparable segmentation and volumetry metrics across experiments. In a literature where reproducibility and cross-study comparison remain ongoing challenges, this kind of standardized evaluation scaffold is an important step toward more systematic assessment of automated tumor volumetry approaches. Future Work If the project were to continue beyond the capstone, the next logical steps would include completing multi-fold nnU-Net training, evaluating fold ensembles, and extending training on higher-end hardware so that the model can reach a more mature stage of

convergence. Additional directions include experimenting with alternative architectures such as transformer-based 3D networks, incorporating multimodal inputs (for example, PET/CT) when available, and validating the pipeline on external datasets to quantify domain shift. Over time, hardening this framework into an open, well-documented benchmark could help other researchers avoid similar infrastructure and training hurdles, and focus more directly on improving the underlying models, accelerating progress toward clinically reliable automated tumor volumetry. Acknowledgments I would like to thank Professor Anna Marbut, PhD Candidate, MS (University of San Diego), for her guidance throughout the development of this capstone project and for her mentorship in shaping the structure, clarity, and methodological rigor of this work. Her feedback and encouragement were essential to translating the technical components of this study into a cohesive and academically grounded research project. I am also deeply grateful to Professor Andrew Van Benschoten, PhD (University of San Diego), whose expertise in machine learning and research strategy has been instrumental in defining the research direction and sustaining the broader line of inquiry that this project contributes to. His mentorship has greatly enriched both the technical foundation and scientific relevance of this work. I extend my sincere thanks to Dr. Gary Takahashi, MD, whose clinical insight and willingness to discuss the practical implications of automated volumetry significantly strengthened the framing of the problem and the interpretation of its clinical significance. His

410