Scale-adaptive ROI and CBAM enhanced MobileNetV2 for low-dose CT pulmonary nodule classification

Abstract

Accurate detection of pulmonary nodules in low-dose computed tomography (LDCT) is crucial for early screening of lung cancer. To address the challenge of classification accuracy degradation caused by multi-scale nodule variations and complex background interference, this paper proposes an enhanced MobileNetV2 model based on scale-adaptive regionof-interest (ROI) cropping and CBAM attention fusion. Specifically, the proposed method first dynamically crops ROI to focus on lesion regions, and then integrates the Convolutional Block Attention Module (CBAM) to reinforce the representation of key features. Experimental results on the LIDC-IDRI dataset demonstrate that the model achieves a high classification accuracy of 99.36% and an Area Under the Curve (AUC) of 0.9993, representing an absolute improvement of 12.18% in accuracy compared to the baseline model. The aforementioned approach provides an efficient and lightweight solution for automatic nodule screening in clinical LDCT scenarios.