Self-DenseMobileNet: A robust framework for lung nodule classification using Self-ONN and stacking-based meta-classifier

Accurate classification of lung nodules can significantly improve treatment outcomes and patient survival rates by enabling early detection and preemptive intervention in cases of suspected lung cancer. Due to the limited availability of sufficiently comprehensive chest X-ray (CXR) datasets, we utilized a composite dataset from the NODE21 challenge to classify nodules and non-nodules in chest radiographs. In this study, we propose a novel computer-aided diagnosis (CAD) framework that includes image standardization and advanced image enhancement techniques for improved quality. We integrated four state-of-the-art deep learning (DL) models such as DenseNet201, MobileViTv2-0.50, MobileViTv2-0.75, and ResNet152, into this framework. We also introduced our own proposed Self-DenseMobileNet, which was trained on four image-enhanced variants of our dataset. The prediction probabilities from this model were transformed into tabular form, and then eight classical machine learning (ML) models were trained. The top three performing classical models were subsequently combined using a stacking algorithm to create a meta-classifier. Finally, We used class activation mapping (CAM) to demonstrate the interpretability of the best-performing model. Our proposed framework achieved an accuracy of 99.28%, a precision of 99.60%, a recall of 99.47%, a specificity of 98.68%, an F1-score of 99.53%, and an area under the curve (AUC) of 0.99 on internal validation images using a Meta-Random Forest Classifier. When tested on a completely different dataset, the external validation results showed an accuracy of 89.40%, a precision of 90.08%, a recall of 92.58%, a specificity of 84.6%, and an F1-score of 91.31%, and an AUC of 0.90. The novel framework resulted in a significant performance enhancement in classifying CXRs with lung nodules. The classification reasoning of the models can also be explained by ScoreCam heatmaps, which further solidifies its robustness. Our proposed model will assist healthcare experts in making informed decisions regarding lung cancer.