AAI_2025_Capstone_Chronicles_Combined

1

Enhancing Lung Tumor Volume Accuracy on CT A 3D Deep Segmentation and Reconstruction Pipeline with Clinical RTSTRUCT Integration

Laurentius von Liechti Master of Science in Applied Artificial Intelligence Shiley Marcos School of Engineering / University of San Diego​ lvonliechti@sandiego.edu

voxel-based tumor volume computation. All preprocessing, model configuration, and data splits are driven by nnU-Net’s self-configuration logic for the 3d_fullres setting, and training is performed on cloud-hosted GPUs using a single fold that was truncated after 263 epochs due to the seven-week capstone timeline and compute constraints. The system produces case-level segmentation metrics (Dice and Jaccard) and absolute and relative volume errors that can be stratified by tumor size and acquisition characteristics. Under these constraints, the trained model demonstrates clear learning behavior but limited generalization on a held-out test set, with low overlap metrics and systematic overestimation of tumor volume. The more durable contribution is the DICOM-native, geometry-preserving pipeline itself, which provides a transparent and reusable template for future work benchmarking segmentation models and investigating the sources of volumetric error in lung tumor CT.

ABSTRACT

Reliable lung tumor volumetry from computed tomography (CT) remains difficult due to inter-observer variability in gross tumor volume (GTV) delineation and sensitivity to acquisition parameters such as slice thickness and reconstruction kernel (Das et al., 2021; Choe et al., 2019). Although deep learning–based segmentation methods have shown promising results, they are rarely evaluated within standardized, DICOM-native workflows that preserve geometric consistency, which complicates fair comparison across models and studies (Wang et al., 2024). This project uses the NSCLC-Radiomics (Lung1) dataset, which provides pretreatment CT scans with clinically defined GTV contours (Aerts et al., 2019), to develop and test an end-to-end, reproducible pipeline for automated tumor volumetry. The proposed framework integrates a 3D nnU-Net v2 segmentation model, strict DICOM geometry handling, RTSTRUCT voxelization into voxel-aligned masks, and standardized

398