AAI_2025_Capstone_Chronicles_Combined

3

prediction in endoscopic imaging (Qadir et al., 2021). However, many of these methods are evaluated in isolation, without a standardized, DICOM-native pipeline for converting segmentation outputs into clinically interpretable tumor volumes. In practice, this means that differences in preprocessing, contour voxelization, and volume calculation can obscure whether performance gains arise from the model itself or from downstream implementation choices. These limitations motivate the need for a geometry-consistent, reproducible volumetry framework that can be used both to support clinical workflows and to provide a common backbone for fair comparison of future segmentation models. Purpose The purpose of this project is to address these gaps by developing a DICOM-native, geometry-consistent pipeline for automated segmentation and tumor volume estimation using lung CT. The central research question is: Can a standardized pipeline built around 3D deep learning segmentation and precise geometric reconstruction produce accurate, reproducible tumor volumes across heterogeneous CT acquisitions? The target end users include radiologists, radiation oncologists, medical physicists, and imaging researchers who require consistent volumetric measurements for clinical workflows or for benchmarking segmentation models. The project uses the NSCLC-Radiomics (Lung1) dataset (Aerts et al., 2019), which includes pre-treatment CT scans paired with clinician-drawn GTV contours encoded in DICOM RTSTRUCT format. These contours serve as ground truth for training and evaluation. In a live deployment environment, the same inputs would be obtained directly from clinical

CT scanners and radiotherapy planning systems. The project’s broader goal is not merely to evaluate nnU-Net or compare segmentation algorithms, but to create a repeatable volumetry framework that integrates: (1)​ standardized DICOM ingestion and spatial normalization,​ (2)​ contour voxelization based on true patient-space geometry,​ (3)​ 3D deep learning segmentation, and​ (4)​ physically accurate voxel-based volume computation. By controlling for geometry, metadata, and preprocessing across all experiments, the framework enables meaningful model comparisons and provides a foundation for future studies on segmentation robustness, radiomics stability, and volumetric error attribution. This work therefore aims to produce not only model performance results, but a reusable template that advances reproducibility in lung tumor volumetry research. 2 Data Summary This project uses the NSCLC-Radiomics (Lung1) dataset, a publicly available thoracic oncology cohort hosted on The Cancer Imaging Archive (TCIA) that provides pre-treatment CT scans and expert-drawn gross tumor volume (GTV) contours for patients with non-small cell lung cancer (Aerts et al., 2019). Each case includes a complete DICOM CT image series and an accompanying DICOM RTSTRUCT file that encodes the clinical GTV using contour polygons defined directly in patient-coordinate space (Wee et al., 2019). The presence of both raw imaging data and radiotherapy-grade contours makes this dataset well suited for voxel-wise segmentation research and volumetric analysis.

400