Classifying non-smokers and smokers on chest X-ray using baseline convolutional neural networks and VGG16

Timely identification of smoking-related lung abnormalities is essential for effective clinical intervention, as smoking is a significant risk factor for a variety of pulmonary diseases. This work proposes a deep learning-based methodology to categorize an individual as a smoker or non-smoker based on chest X-rays (CXR). Utilizing the capabilities of Convolutional Neural Networks (CNNs), the proposed model examined the efficacy of a baseline CNN architecture and a transfer learning model derived from VGG16. A curated dataset of 5856 labeled chest X-ray images was utilized, and random shuffling and augmentation were used to oversample the dataset to address the data imbalance issue and then preprocess to train and assess both models. The baseline CNN achieved a classification AUC score of 95.15%, illustrating the efficacy of end-to-end learning in retrieving radiographic characteristics associated with smoking. The VGG16-based model, fine-tuned for classification, significantly surpassed the baseline with an accuracy of 98.74%, demonstrating that deep feature representations acquired from extensive datasets can be efficiently used for medical imaging tasks. These results highlight the feasibility of using CNNs for automated, noninvasive screening of smoking history from CXR images. The proposed method can assist in making correct clinical decisions, particularly where patient history is incomplete or unreliable. Furthermore, this approach has potential applications in public health surveillance and early risk assessment, especially in resource-limited settings like primary healthcare centers in rural areas.