AAI_2025_Capstone_Chronicles_Combined

4

The dataset includes detailed imaging metadata - such as PixelSpacing, slice thickness, ImagePositionPatient/Orientation vectors, and scanner-specific acquisition parameters - which enables accurate mapping of voxel indices into physical space. This geometric information is essential for reconstructing 3D masks from the RTSTRUCT contours using a consistent reference frame, and it allows tumor volumes to be computed in true patient-space units rather than relying on approximations or assumptions about slice spacing. The GTVs were delineated by clinical experts as part of radiotherapy treatment planning, providing a clinically meaningful and widely adopted reference standard for evaluating segmentation algorithms (Van De Steene et al., 2002; Das et al., 2021). Variables and Data Types The primary imaging data consist of 3D CT volumes, stored as voxel intensities in Hounsfield Units. These serve as the input to the segmentation models. The RTSTRUCT files

contain lists of closed polygon contours for each axial slice where the tumor is present, represented in millimeter-accurate patient coordinates. From these contours, a voxelized 3D binary mask is reconstructed, aligned precisely to the CT grid; this mask constitutes the supervised training target and the baseline for evaluating segmentation accuracy and derived tumor volumes. Spatial metadata such as PixelSpacing (Δx, Δy) and slice spacing Δz - derived from ImagePositionPatient and ImageOrientationPatient - provide the necessary factors for converting voxel counts into physical volumes. Acquisition-related categorical variables (e.g., reconstruction kernel, scanner vendor) are also available and will support stratified analyses examining how protocol differences influence segmentation and volumetry performance (Choe et al., 2019; Wang et al., 2024). A detailed summary of all variables and their roles in the project is provided in Table 2.1.

Table 2.1. Variables and Data Types Used in This Project Data Component Type Description

Use in Project

CT voxel intensities

3D float tensor (HU)

Raw CT imaging volume

Model input for segmentation and volume inference

RTSTRUCT GTV contours

Polygon coordinate lists (mm)

Clinically delineated tumor boundaries

Ground-truth reference for segmentation

Reconstructed GTV mask PixelSpacing (Δx, Δy)

3D binary tensor (z, y, x)

Voxelized tumor region aligned to CT

Supervised training target; evaluation baseline Needed for physical volume computation

Float pair

In-plane resolution (mm)

401