Lung cancer remains one of the most prevalent cancers worldwide and has a low overall five-year survival rate, making early detection a critical clinical priority. The advent of AI-based Computer-Aided Detection (CAD) systems for lung cancer has significantly reduced the workload of radiologists. Traditionally, these AI models are trained on curated datasets that are extensively annotated and reviewed by multiple radiologists. In modern clinical workflows, radiologists frequently interact with AI-generated results to aid interpretation and streamline reading and reporting processes. These routine interactions offer a valuable opportunity to enhance AI algorithms further. However, regulatory frameworks such as HIPAA and GDPR restrict the centralized sharing of patient data due to privacy concerns, limiting the ability to aggregate and retrain models centrally. To overcome this challenge, we adopt a decentralized learning framework that preserves patient privacy. Our approach leverages data generated during routine interpretation of thoracic CT scans to retrain an existing lung nodule detection model. We propose an AI/ML-based methodology designed to address the inherent noise in annotations generated during routine clinical readings. To further improve efficiency, we incorporate a negative mining strategy during retraining, which reduces computational overhead and enables faster model updates. We evaluate the retrained model on a Lung Cancer Screening dataset (n = 1128). Compared to the baseline AI model, our decentralized model achieves improvements in sensitivity of 6.90%, 6.98%, 5.65%, 4.05%, 3.55%, 3.87%, and 3.32% at 1/8, 1/4, 1/2, 1, 2, 4, and 8 false positives per scan, respectively. These results demonstrate a scalable, efficient, and privacy-preserving approach to continuously improve AI algorithms using data generated during routine clinical practice.