Small but Strong: Lightweight Architecture Improves Lung Nodule Detection and Segmentation

Lung nodule detection and segmentation from computed tomography (CT) scans are essential for lung cancer diagnosis and therapy. Inspired by the substantial progress of neural networks in 2D natural images, current methods often adapt 2D networks to 3D for these tasks, resulting in massive network parameters. However, the limited training data available for lung nodule CT scans (≈2 k) compared to natural images (>10 b) lead to overfitting and reduced performance. This study proposes a lightweight architecture for lung nodule detection and segmentation, alleviating the overfitting caused by the data scarcity. To reduce the parameters of the backbone network, the deformable Transformer is introduced to retain large receptive fields for desired performance with a small model size. To eliminate extensive anchor‐related parameters, this study employs the anchor‐free structure for the detection task, which also aligns the common features with the segmentation task. Benefiting from this, a higher level of parameter sharing becomes possible between two tasks, further reducing the entire model size and imposing a strong regularizer to alleviate the overfitting. The proposed method is validated on the LUNA16 and LIDC‐IDRI datasets, and experimental results demonstrate that the method outperforms the state‐of‐the‐art methods with a small model size. Code: https://github.com/LiuLiluZJU/LWNet.