NRDPNet: Noise-Resistant Dual-Path Residual Network for High-Precision Lung Nodule Detection
TL;DR
A novel 3D deep neural network (DNN) architecture that employs an encoder–decoder framework and integrates a region proposal network to achieve precise localization of candidate nodules is proposed, validating its effectiveness and generalizability within simulated image degradation conditions.
Abstract
Lung nodule detection holds significant clinical importance in low-dose computed tomography (LDCT) lung cancer screening. However, during actual imaging, issues such as quantum noise amplification and diminished image quality may compromise detection algorithm stability. To address these challenges, this paper proposes a novel 3D deep neural network (DNN) architecture, NRDPNet. This model employs an encoder–decoder framework and integrates a region proposal network (RPN) to achieve precise localization of candidate nodules. At the core, we propose a TiedSE-GCT Multi-Scale Residual Block (TG-MSRB), combined with a Tied Block Squeeze-and-Excitation (TiedSE) module and a Gated Channel Transform (GCT) module, enhancing the model’s feature representation capability and interference resistance. To systematically evaluate the model’s stability under low image quality conditions, this paper constructs noise perturbation environments of varying intensities based on a Poisson statistical model, simulating imaging scenarios with increased image noise. Experimental results on the LUNA16 dataset demonstrate that the proposed method exhibits minimal performance fluctuation across varying noise intensities, showcasing robust stability. Concurrently, under standard evaluation metrics, the model outperforms multiple mainstream detection methods in terms of detection sensitivity and competition performance metric (CPM), validating its effectiveness and generalizability within simulated image degradation conditions.