ARTIFICIAL INTELLIGENCE-DRIVEN PULMONARY NODULE IDENTIFICATION VIA COMPUTED TOMOGRAPHY: TECHNOLOGICAL PROGRESS, UNRESOLVED HURDLES, AND EMERGING HORIZONS

Abstract

Globally, lung malignancy stands as the foremost oncological cause of death, responsible for roughly 2.5 million incident cases and 1.8 million fatalities recorded in 2022. Stage-specific prognosis diverges sharply: patients whose tumors are identified at an early, localized phase achieve five-year survival exceeding 80%, whereas those diagnosed at stage III or IV face outcomes below 10%. Despite low-dose computed tomography (LDCT) having been adopted as the recommended modality for population-level screening programs, reliably identifying sub-centimeter pulmonary nodules remains technically demanding, and interpretation variability across radiologists constitutes a persistent clinical problem. This systematic review evaluates the contribution of deep learning (DL) to automating nodule detection, volumetric segmentation, and malignancy risk stratification from CT examinations. Through a structured search of PubMed, Scopus, and IEEE Xplore spanning 2018 through 2025, we synthesize evidence on convolutional architectures, three-dimensional encoder–decoder networks, attention-augmented models, and vision transformer frameworks. Leading systems attain sensitivity figures between 92% and 96%, with Competition Performance Metric (CPM) values surpassing 0.90 on the LUNA16 benchmark. Generative adversarial approaches have shown promise in mitigating training-data scarcity, while gradient-weighted class activation mapping (Grad-CAM) has emerged as an important transparency tool for clinical adoption. Although DL markedly elevates both throughput and diagnostic fidelity, persistent barriers, including cross-scanner generalization, residual false-positive burden, limited model interpretability, and regulatory compliance demands, must be addressed before routine clinical integration can be realized.