Systematic Evaluation of 2D and 3D Deep Learning Detectors for Pulmonary Nodule Detection in Chest CT

Abstract

Accurate detection of pulmonary nodules in chest CT is essential for early lung cancer diagnosis. This study presents a systematic evaluation of two-dimensional (2D) and three-dimensional (3D) deep learning detectors for pulmonary nodule detection using the LUNA16 benchmark dataset. While 3D deep learning approaches can exploit full volumetric information, 2D detectors remain attractive due to their lower computational complexity and simpler training requirements. Particular attention was devoted to the effect of annotation preprocessing in the 2D setting. Two 2D detectors (RetinaNet and YOLO) and two 3D detectors (3D RetinaNet and a hybrid CNN-Transformer architecture) were evaluated. For the 2D experiments, multiple dataset variants were prepared by filtering annotations according to minimum bounding-box area and by balancing positive and negative slices. Detection performance was
evaluated using F1-score, precision, recall, mAP, mAR, and FROC-based metrics across multiple IoU thresholds. The results showed that preprocessing strategies substantially affected 2D detector performance. Filtering very small peripheral annotations and balancing the datasets improved localisation quality and detector stability for both RetinaNet and YOLO. Among the 2D configurations, the best results were obtained for the balanced filtered subsets. The 3D RetinaNet achieved the strongest overall performance, providing higher detection sensitivity, more stable localisation, and better generalisation between validation and independent test data. The CNN-Transformer also benefited from volumetric input and retained relatively high recall at lower IoU thresholds, although its performance
remained lower than that of the 3D RetinaNet. The findings demonstrate the importance of volumetric context for robust pulmonary nodule detection and indicate that appropriate annotation preprocessing can substantially improve the effectiveness of slice-based 2D detection frameworks.