LNA-Net: Enhancing detection and classification of benign and malignant pulmonary nodules in CT scans

Lung cancer remains a leading cause of cancer-related mortality worldwide, highlighting the critical need for accurate and efficient detection methods. CT scans involve multiple layers of imaging, requiring meticulous examination to identify lung nodules. Traditional manual methods and many AI-based detection approaches are often slow and computationally intensive. In this paper, we present LNA-Net, a novel lightweight deep learning framework designed for real-time detection and streamlined classification of benign and malignant pulmonary nodules in CT scans. We introduce the novel Adaptive Kernel Attention Module (AKAM), a plug-and-play attention block that enhances the extraction of features for varying-size and small nodules, along with a customized C2fCNB block to improve feature fusion in the neck of the network. By integrating nodule detection and malignancy classification into a unified pipeline, LNA-Net eliminates the need for an additional classifier. Evaluated on the LUNA16 dataset, our model outperformed state-of-the-art (SOTA) object detection frameworks, achieving 82.5% mAP@50 for nodule detection, representing a 3.4% improvement over YOLOv10n, along with 93.4% recall and 93.3% F1-score in classifying nodules, a 9.4% and 5.4% enhancement compared to the YOLOv10 backbone respectively. LNA-Net’s high accuracy and real-time efficiency position it as a valuable tool for assisting radiologists in the timely diagnosis of lung cancer. • Proposed Adaptive Kernel Attention Module for better multi-scale feature extraction. • Proposed C2fCNB block for feature fusion enhancement. • Designed a unified framework for nodule detection and malignancy classification.