JMIR Publications

ABSTRACT

Background:

Lung cancer remains the leading cause of cancer-related mortality in Taiwan, with pulmonary nodules serving as critical early indicators. Early detection through computed tomography (CT) imaging significantly improves patient survival rates. Deep learning approaches, particularly the You Only Look Once (YOLO) family of object detection algorithms, have shown promise in automating nodule detection, potentially reducing diagnostic workload and improving screening sensitivity.

Objective:

This study aimed to evaluate whether integrating Residual Network (ResNet) architecture into the YOLOv5 backbone improves pulmonary nodule detection performance compared to standard YOLOv5 models, and to assess the impact of contrast enhancement preprocessing on detection accuracy.

Methods:

We utilized the LUNA16 dataset, comprising 888 CT scans with 1,186 expert-annotated nodules (≥3mm diameter). Three-dimensional CT images were converted to 2D slices, with optional contrast-limited adaptive histogram equalization (CLAHE) preprocessing. Four model architectures were compared: YOLOv5m, YOLOv5l, ResNet-YOLOv5m, and ResNet-YOLOv5l, each trained across four hyperparameter configurations. Performance was evaluated using precision, recall, and mean average precision (mAP) on a held-out test set (n=238 images).

Results:

ResNet-enhanced models demonstrated superior performance, particularly in recall—a critical metric for medical screening. On original CT images, ResNet-YOLOv5l achieved the highest mAP (76.75% ± 1.34) and recall (71.78% ± 2.08) compared to YOLOv5l (mAP: 73.75% ± 4.42; recall: 69.53% ± 6.55). With CLAHE preprocessing, ResNet-YOLOv5l showed further improvement (mAP: 74.95% ± 0.84; recall: 72.78% ± 2.67), outperforming all standard YOLOv5 variants. Training time differences remained within one hour between comparable architectures.

Conclusions:

Integrating ResNet residual blocks into YOLOv5 significantly enhances pulmonary nodule detection performance, with improved recall reducing potential missed diagnoses. CLAHE preprocessing provides additional performance gains. These findings support the clinical applicability of ResNet-enhanced YOLO architectures for computer-aided lung cancer screening.