Author: Xuezhen Feng, Li-Sheng Geng, Haoze Li, Xi Liu, Tianyu Xiong, Ruijie Yang 👨🔬
Affiliation: Department of Health Technology and Informatics, The Hong Kong Polytechnic University, School of Physics, Beihang University, School of Nuclear Science and Technology, University of South China, Department of Radiation Oncology, Peking University Third Hospital 🌍
Purpose: This study aimed to develop a deep learning-based algorithm for automatically delineate gross tumor volume (GTV) for lung cancer patients, alleviating the workload of radiologists and improving the homogeneity of radiotherapy.
Methods: A public dataset with 422 non-small cell lung cancer patients was used for algorithm training, validation, and testing. Each patient had access to pretreatment CT scans, along with label of the GTV. Image preprocessing steps, encompassing intensity clipping, cropping, and normalization were implemented to emphasize pertinent information and reduce computational demands. We developed an algorithm for segmenting the GTV based on a 3D encoder-decoder architecture, incorporating a large kernel attention mechanism into its structure to enhance its capability of capturing long-range dependencies and learning relationships between different parts, in contrast to using a conventional attention gate mechanism. We comprehensively evaluated the algorithm performance using precision, sensitivity, specificity, Dice similarity coefficient (DSC), relative absolute volume difference (RAVD), and average symmetric surface distance (ASSD).
Results: Our algorithm achieved promising results on the independent test dataset, with an average DSC of 0.752, an ASSD of 3.01 mm, a RAVD of 0.372 mm, a precision of 0.815, a specificity of 0.998, and a sensitivity of 0.756. When compared to the algorithm with the same architecture but conventional attention gate mechanism, our model demonstrated significant improvements: 18.87% for DSC, 58.42% for ASSD, 78.30% for RAVD, 12.29% for precision, and 16.94% for sensitivity. All p-values were below 0.05, indicating that our model significantly outperformed the one with the conventional attention gate mechanism across all evaluation metrics.
Conclusion: We developed a deep learning algorithm incorporating a novel attention mechanism for contouring the GTV on thoracic CT images. Compared with conventional attention gate mechanism, our algorithm exhibited superior performance. To further demonstrate its effectiveness and robustness, external validation is necessary.