AAI_2025_Capstone_Chronicles_Combined

5

Slice spacing (Δz)

Float (derived)

Distance between slices from DICOM geometry

Needed for physical volume computation Stratified evaluation (e.g., error vs. slice thickness)

Acquisition metadata Categorical

Reconstruction kernel, scanner vendor, etc.

Data Quality and Integrity Checks Because RTSTRUCT contours are defined in patient-space rather than slice-indexed voxel space, accurate reconstruction of the GTV mask requires strict alignment between the CT and RTSTRUCT objects. A series of validation steps were performed, including: matching CT and RTSTRUCT objects via FrameOfReferenceUID, verifying slice order using the geometric coordinates encoded in ImagePositionPatient, and ensuring that no orientation flips or axis permutations were introduced during preprocessing. All volumetric data transformations were anchored to the image geometry provided by SimpleITK, which preserves the DICOM origin, direction cosines, and spacing throughout the pipeline. This approach eliminated the most common source of error in mask reconstruction - orientation mismatch between the CT array and the voxelized contour representation. Exploratory Observations Initial data exploration revealed that in-plane voxel spacing is typically in the 0.9–1.2 mm range, with slice spacing commonly between 2 and 3 mm, consistent with standard thoracic CT protocols. Tumor volumes vary substantially across patients, from very small lesions (<1 mL) to large, heterogeneous masses exceeding 100 mL. This variability has methodological implications: small tumors amplify boundary-related segmentation errors, while large or highly irregular tumors challenge a model’s ability to capture long-range 3D

structure and surface topology. Tumor morphology also spans a wide spectrum - from compact nodules to infiltrative, spiculated masses with non-tissue internal volumes (Figure 2.1) - providing a rigorous basis for evaluating model generalization across shape complexity (Kashyap et al., 2025; Hiraman et al., 2024). Relationship to Project Goals The reconstructed GTV mask is the quantitative ground truth that the model is expected to reproduce. The CT intensity data form the input to the segmentation network, and the voxelized mask provides the supervised training labels. Tumor volume is computed from the predicted mask using a straightforward DICOM-consistent formula: which is then converted to milliliters. Agreement between predicted and ground-truth volumes will be assessed using relative volume error and stratified analyses examining sensitivity to acquisition parameters such as slice thickness and reconstruction kernel. This approach mirrors prior work demonstrating that volumetry is highly sensitive to segmentation boundary fidelity and CT acquisition characteristics (Yoon et al., 2022; Choe et al., 2019).

402