2408 - Striking the Balance in Learning: Precision and Context in Lung Tumor Segmentation with 3D UNet Models

S. Sarkar¹, Y. Miao¹, P. T. Teo², and M. Abazeed³; ¹Northwestern University, Chicago, IL, ²Department of Radiation Oncology, Northwestern University, Feinberg School of Medicine, Chicago, IL, ³Department of Radiation Oncology, Northwestern University Feinberg School of Medicine, Chicago, IL

Purpose/Objective(s): This study aims to determine if utilizing 3D UNet models, trained and tested on diverse datasets, can improve lung tumor segmentation accuracy and sensitivity. Additionally, it investigates the impact of varying region-of-interest (ROI) sizes on learning and inference processes to achieve an optimal balance between precision and contextual understanding in lung tumor segmentation. Materials/

Methods: A meticulously designed methodology was employed for lung GTV segmentation, centered around the implementation of a sophisticated 3D UNet architecture. Drawing upon datasets from two academic institutions, encompassing a total of 941 cases, the study ensured a diverse and representative sample. Through meticulous data sampling techniques, a balance was achieved between full CT scans (512x512x256) and context windows (128x128x128) centered around tumor isocenters. Both training and testing phases were conducted comprehensively, encompassing analysis on entire CT volumes and specific ROIs. The computational backbone of the research was supported by distributed computation, enabled by two Nvidia Tesla A100 40GB GPUs, facilitating efficient model training. With a batch size of 12, the model underwent 20 epochs, concluding within approximately 12 hours.
Results: The study yields significant insights into lung tumor segmentation accuracy and contextual understanding. When trained and tested exclusively on localized ROIs, the model demonstrates commendable performance with a DICE score of 0.576, indicating substantial precision within confined areas. However, the transition to the full CT context unveils a notable decrease in segmentation accuracy, with the DICE score dropping to 0.3002, underscoring the inherent trade-off between precision and contextuality. Furthermore, performance disparities between the contributing institutions are observed, with a noteworthy 26.04% gap in DICE scores. The utilization of a context window approach offers exceptional precision in local segmentation but sacrifices the ability to capture the broader tumor context.
Conclusion: The study underscores the critical importance of achieving a delicate balance between precision and context in lung tumor segmentation. While training exclusively on ROIs yields satisfactory results, it falls short of capturing the comprehensive tumor landscape. Conversely, leveraging the full CT context provides valuable insights but at the expense of precision. To address this inherent challenge, a hybrid model is proposed, integrating the strengths of both approaches. By harnessing the full CT model for identifying regions of interest and utilizing the context window for precise segmentation, this synergistic approach not only surpasses human precision but equips clinicians with a robust tool for accurate and comprehensive tumor delineation. Such advancements hold immense potential for enhancing treatment planning efficacy and ultimately improving patient outcomes in clinical settings.