Abstract

This study systematically reviews the past five years of convolutional neural network (CNN) research in pulmonary imaging for screening, triage, and follow-up. Guided by PRISMA, we analyzed English-language studies (2020–2025) from PubMed/MEDLINE, IEEE Xplore, Scopus, and Google Scholar, focusing on CT/LDCT and chest X-ray (CXR) applications for detection, segmentation, and prognosis. Data extraction was standardized across datasets, preprocessing, model architectures, validation strategies, and evaluation metrics. Results reveal a convergent pipeline of detection → segmentation → quantification → decision support. On CT, 2.5D/3D candidate generation combined with boundary-aware segmentation improves performance for small nodules and ground-glass opacities. On CXR, integrating global and regional features with anatomical priors (e.g., bone suppression) mitigates projection overlap. Weak, semi-, and self-supervised learning, along with contrastive learning and knowledge distillation, enhance robustness under limited data and domain shift, while focal-type losses address class imbalance. Deployment-oriented optimizations (e.g., ONNX, TensorRT, pruning, and quantization) significantly reduce inference latency and facilitate integration with clinical systems (PACS/RIS) via structured outputs and saliency visualization. Strengths include clinically aligned pipelines and improved efficiency, whereas limitations persist in external validation, calibration, and reporting transparency. We recommend routine external “test-only” evaluation, prospective validation, standardized uncertainty reporting, and improved reproducibility practices. These steps are essential to advance CNN-based pulmonary imaging systems from experimental feasibility toward reliable clinical deployment.

Keywords