A UNIFIED FRAMEWORK WITH SPATIO-TEMPORAL AND MULTI-SCALE FEATURE LEARNING FOR 3D LUNG NODULE DETECTION

Abstract

Lung nodule detection in CT scans remains a challenging computer vision task due to structural ambiguity and limited annotations. This paper introduces I3DR-Net, a one-stage 3D deep learning framework for joint detection and classification. The model combines a pre-trained I3D backbone for spatio-temporal feature extraction with a modified FPN for enhanced multi-scale representation. Experimental results show competitive performance (≈70–73% accuracy), demonstrating that transfer learning and improved pyramid design effectively enhance detection robustness for computer-aided diagnosis applications.