2023 North America Conference on Lung Cancer (NACLC) - Posters and Abstracts-PP01.132 Yaqi Miao NACLC23 Abstract

PP01.132 Yaqi Miao NACLC23 Abstract

Pdf Summary

Tumor segmentation plays a crucial role in radiation tumor treatments. However, manual segmentation can introduce variability and uncertainty. This study explores the use of deep learning engines to improve accuracy and precision in tumor delineation, potentially mitigating treatment toxicity and enhancing efficacy.<br /><br />To address the issue of operator-dependent uncertainties in ground truth labeling, the researchers used contours associated with recurrence-free treatment outcomes as a selection criterion for training their deep learning segmentation model. The trained model was then applied to a dataset with local recurrence, and the differences between clinician and deep learning-generated contours were statistically assessed.<br /><br />Multiple U-Net configurations, including 2D U-Net, Stacked U-Net, and Ipsilateral Stacked U-Net, were developed and trained on a dataset comprising CT scans from two clinical organizations. Transfer learning, data augmentation, and post-processing techniques were employed to enhance model performance. The accuracy and precision of the models were measured using metrics such as Jaccard Index (JI), Mean Squared Error (MSE), and Dice Similarity Coefficient (DSC).<br /><br />Preliminary findings revealed that the Ipsilateral Stacked U-Net outperformed other engines, achieving a JI of 0.3316, MSE of 0.0019, and DSC of 0.4925. Qualitative assessments supported these results, indicating that this model closely replicated manual contour delineation, suggesting a potential reduction in segmentation variability.<br /><br />The successful implementation of the 2D and Stacked U-Net models demonstrates promising avenues for improving tumor delineation accuracy and precision. The achieved DSC of 0.49 surpasses the performance reported in current publications, highlighting the potential of this approach to automate tumor contour delineation realistically. Ultimately, this advancement could contribute to improved treatment outcomes for lung cancer patients.<br /><br />Keywords: Automatic Lung Tumor, Auto-Segmentation, Deep Learning, U-Net, Radiation Oncology, CT Scan, Data Augmentation.

Keywords

Tumor segmentation

deep learning engines

accuracy

precision

tumor delineation

ground truth labeling

U-Net configurations

Ipsilateral Stacked U-Net

CT scans

lung cancer patients