<p>Accurate detection of lung nodules in chest radiographs plays a critical role in the early diagnosis of pulmonary diseases, including lung cancer. In this study, we integrate deep learning techniques for nodule classification with object detection approaches. For model training and evaluation, we utilized the publicly available ChestX-ray14 dataset, which comprises approximately 112,000 chest X-ray images, including 6,371 labeled as containing lung nodules. We implemented several pretrained convolutional neural network (CNN) architectures—specifically DenseNet121, ResNet34, AlexNet, and InceptionNet—for binary classification tasks aimed at identifying the presence of lung nodules. To gain interpretability into the decision-making processes of these models, we employed Gradient-weighted Class Activation Mapping (Grad-CAM). In addition to classification, we trained an object detection model designed to identify and localize nodules by generating bounding boxes around them. We assessed both the classification and detection models by comparing their outputs and evaluating their performance, with the goal of better understanding their effectiveness in automated lung nodule detection. Overall, these findings demonstrate that transfer learning based classification models achieve high sensitivity but suffer from low precision on weakly labeled data, while object detection models provide more reliable localization when trained on expert annotated datasets. The results highlight the critical impact of annotation quality and suggest that detection based approaches may be better suited for clinically meaningful lung nodule identification. These findings also show that while ensemble classification improves sensitivity and ROC-AUC on weakly labeled data, object detection trained on expert annotations yields substantially higher precision and more clinically reliable localization.</p>