2023 North America Conference on Lung Cancer (NACLC) - Posters and Abstracts-PP01.132 (Poster) Auto-Segmentation of Lung Tumors Using Deep Learning Engines

Pdf Summary

Doctors manually identifying tumor areas in CT scans for radiation therapy can introduce variance in treatment outcomes. Deep learning engines can potentially improve tumor delineation by increasing accuracy and precision. In this study, researchers developed and trained multiple U-Net configurations, including 2D U-Net, Stacked U-Net, and Ipsilateral Stacked U-Net, on a dataset of CT scans from 1069 patients. Techniques such as transfer learning, data augmentation, and post-processing were employed to enhance model performance. Model accuracy and precision were quantified using metrics like Jaccard Index, Mean Squared Error, and Dice Similarity Coefficient. The trained model was then applied to a dataset with local recurrence to compare with clinician-generated contours.<br /><br />The results showed promising avenues for enhancing the accuracy and precision of tumor delineation using 2D and Stacked U-Net models. The Dice Similarity Coefficient of 0.51 surpassed the performance reported in current publications, indicating the potential for realistic automation of tumor contour delineation. This can contribute to improved lung cancer treatment outcomes.<br /><br />The 2D U-Net model, initialized with pre-trained weights for detecting abnormality in brain MRI, showed good performance on CT slices and corresponding Planning Target Volumes (PTVs). The Stacked U-Net model, which captures more local context, required training from scratch but achieved high performance. The Ipsilateral Stacked U-Net model, trained on the ipsilateral lung to remove unnecessary information, performed even better in testing.<br /><br />The study aimed to reduce operator-dependent uncertainties in ground truth labeling by utilizing contours associated with recurrence-free treatment outcomes as a selection criterion for training the deep learning segmentation model. By decreasing segmentation variability, treatment outcomes for lung cancer patients can potentially be improved. The qualitative assessments also reflected the proficiency of the models in closely replicating manual contour delineation results.<br /><br />Overall, this study showcases the potential of deep learning engines in automating tumor segmentation for improved lung cancer treatment outcomes.

Asset Subtitle

Yaqi Miao

Keywords

tumor delineation

radiation therapy

deep learning engines

U-Net configurations

CT scans

transfer learning

data augmentation

Dice Similarity Coefficient

lung cancer treatment outcomes

ground truth labeling