AAI_2025_Capstone_Chronicles_Combined

7

acquisition-related variation and grounded in accurate geometric representation. Deep Learning Approaches Deep learning has increasingly become the dominant approach for automated lung tumor segmentation on CT. Early 3D convolutional neural network (CNN) architectures, such as 3D U-Net combined with conditional random field refinement, demonstrated improvements in nodule segmentation by leveraging volumetric spatial information unavailable to 2D slice-wise models (Wu et al. 2020). Subsequent approaches introduced more advanced designs tailored to the complex morphology of non-small cell lung cancer (NSCLC). These include multi-resolution frameworks such as Co-ReTr, which enhance contextual representation at multiple scales (Kunkyab et al. 2024), and transformer-based models such as SALM, which incorporate both 2D and 3D attention mechanisms to better capture cross-slice relationships (Gayap & Akhloufi 2025). Multi-institutional evaluations of deep learning–based lung tumor segmentation indicate that such models can reach clinically meaningful performance; however, performance degradation has been observed when models are applied to new scanners or patient cohorts without domain adaptation, demonstrating the practical challenge of generalization (Kashyap et al. 2025). Technically, the methods most relevant to this project rely on 3D CNN architectures that operate directly on volumetric CT data. By applying convolutional filters across three spatial dimensions, these models capture cross-slice continuity and volumetric morphology that cannot be learned by 2D networks. Architectures such as 3D U-Net and

nnU-Net (Isensee et al., 2021) follow an encoder–decoder structure: the encoder progressively abstracts spatial information through layers of 3D convolutions and nonlinear activations, while the decoder reconstructs voxel-wise predictions by fusing coarse semantic features with fine spatial detail via skip connections. nnU-Net builds on this template by automatically configuring hyperparameters - including patch size, feature channel counts, data augmentations, and training schedules - based on dataset-specific characteristics, yielding strong performance across diverse medical imaging tasks without manual architectural tuning. This combination of volumetric context and automated configuration makes these frameworks well suited for delineating lung tumors with irregular shapes and heterogeneous appearance on CT. Generalizability across clinical environments remains a central theme in the segmentation literature. A study evaluating a deep learning–based GTV segmentation model before and after transfer learning at a new institution demonstrated a marked decrease in accuracy without adaptation, with performance substantially recovering once site-specific fine-tuning was applied (Kulkarni et al. 2023). These findings reinforce that, while deep learning models can achieve high accuracy, their robustness under changes in acquisition protocols or patient populations is not guaranteed. Broader evaluations reflect similar trends: a recent meta-analysis found that many deep learning approaches report strong segmentation results but often lack thorough external validation or detailed volumetric error reporting (Wang et al. 2024). A comprehensive review of lung tumor segmentation approaches likewise identified data scarcity, annotation noise, and protocol variability as continuing

404