2024 World Conference on Lung Cancer (WCLC) - Posters-P4.04C.11 Pixel-Wise Pulmonary Nodule Growth Prediction on Low-Dose Computed Tomography with 3D-ConvLSTM Deep Neural Network

P4.04C.11 Pixel-Wise Pulmonary Nodule Growth Prediction on Low-Dose Computed Tomography with 3D-ConvLSTM Deep Neural Network

Back to course

Pdf Summary

The study titled "Pixel-Wise Pulmonary Nodule Growth Prediction on Low-Dose Computed Tomography with 3D-ConvLSTM Deep Neural Network" addresses the crucial role of low-dose computed tomography (LDCT) in the early detection of lung cancer. This early detection is essential for extending patient survival time. The research aims to develop an AI model capable of predicting lung cancer growth at its early stages using LDCT-based follow-up scans.<br /><br />The proposed model utilizes a 3D-ConvLSTM deep neural network to predict the voxel-level growth pattern of pulmonary nodules, providing detailed insights into the nodule's development over time. This prediction assists clinicians in determining earlier treatment plans which can help reduce the risk of malignant nodule development and patient anxiety. The study emphasizes the significance of extracting the post-growth diameter from predicted nodule masks, a metric critical for guiding treatment decisions according to clinical guidelines.<br /><br />The research was conducted using an in-house dataset comprising 299 patients with 1008 scans and 2603 groups of nodule follow-up images. From this, the training dataset included 1343 groups, while the validation dataset contained 576 groups. An additional testing dataset of 198 groups was also utilized. Results demonstrated the superiority of the proposed model over existing methods like 3D ResUNet and GM-AE, particularly in predicting quickly growing nodules.<br /><br />The model achieved higher accuracy metrics: Dice score, Intersection over Union (IoU), and Mean Absolute Error (MAE), reflecting its effectiveness in capturing the growth trajectory of pulmonary nodules. These advancements promise to significantly aid precision cancer diagnosis and treatment planning. This work was a collaborative effort between researchers from Tsinghua University, The First Affiliated Hospital of Xiamen University, and Zhuhai Sanmed Biotech Ltd.

Asset Subtitle

Xing Lu

Meta Tag

Speaker Xing Lu

Topic Screening & Early Detection

Keywords

pulmonary nodule

3D-ConvLSTM

low-dose computed tomography

lung cancer detection

AI model

voxel-level prediction

nodule growth

precision cancer diagnosis

Dice score

clinical guidelines