A Multiscale Anchor-Free 3D CNN Integrated with Denoising for Early Lung Nodule Detection in Low-Dose CT Images

The utilization of computer-aided detection for lung nodules in three-dimensional CT scans, leveraging AI deep learning methodologies, is of paramount importance for the automation and enhancement of detection, which in turn significantly improves the efficacy of subsequent treatment protocols. Low-dose CT (LDCT) has emerged as a prevalent screening modality in clinical settings, due to its reduced radiation exposure and the consequent abundance of available data samples. Despite these advantages, LDCT images are inherently limited by diminished resolution and elevated noise levels, which present obstacles to the robustness of AI-facilitated early lung nodule detection. To address these challenges, this paper introduces an innovative framework that integrates sophisticated denoising techniques with a state-of-the-art 3D convolutional neural network (CNN) that incorporates a multiscale anchor-free paradigm. This paradigm is entirely data-driven, therefore obviating the necessity for priori knowledge in setting bounding boxes, which is required for anchor-based approaches. The study develops three distinct denoising methodologies - namely, the Median Filter, Gaussian Filter, and Contrast Limited Adaptive Histogram Equalization (CLAHE) - and ultimately selects the former two based on their superior performance metrics by the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). These selected methods are subsequently integrated into the training of the multiscale anchor-free 3D CNN, compared with a baseline model that utilizes non-denoised data. The standard evaluation of FROC analysis indicates that the proposed models outperform the majority of existing anchorbased networks, with the model incorporating the Median filter achieving a sensitivity of 0.735 at a false positive rate of 0.125 per scan, thereby signifying an approximate 3-4 % enhancement in predictive capability relative to the baseline model. The contributions of this study are manifold: firstly, the pioneering integration of Median and Gaussian Filter denoising techniques into the 3D CNN framework, which substantially elevates the quality of images and the signal-to-noise ratio, consequently boosting the model's proficiency in detecting lung nodules within LDCT images; secondly, the development of a multiscale anchorfree 3D CNN architecture that facilitates the precise and resilient identification of lung nodules across a spectrum of sizes, devoid of reliance on preconceived anchor boxes; and thirdly, the superiority of the proposed models in terms of sensitivity and predictive power, especially at reduced false positive rates, a factor of critical importance in clinical context, thus potentially augmenting the clinical usage of LDCT screening procedures.