Optimized YOLOv11 model for lung nodule detection

Abstract

Objectives

This study proposes an advanced YOLOv11-based lung nodule detection algorithm that balances high accuracy with efficient computation, addressing the critical need for accurate and timely early diagnosis of lung cancer.

Methods

We replaced the traditional backbone with MobileNetV4, which employs reversible connections to prevent information loss and enhance feature representation, thereby improving the model’s efficiency in processing high-resolution CT scans. We developed a novel C2PSA module, C2PSA-MSDA, which integrates Multi-Scale Dilation Attention (MSDA) to capture multi-scale features more effectively. For the neck part, we introduced the new FreqFusion-BiFPN to enhance feature integration and boundary clarity, thereby reducing false positives. Additionally, we created a new C3k2 module, DyC3k2, to optimize feature fusion. We adopted Focal-inv-IoU for bounding box regression and Slide Loss for classification, which help the model focus more on high-quality predictions while still considering lower-quality ones, leading to more balanced and accurate detection.

Results

Extensive experiments on the LUNAR16 dataset and a proprietary dataset demonstrated significant improvements: precision increased by 4.15 %, recall by 3.23 %, mAP50 by 4.04 %, and mAP50-95 by 3.28 % compared to the baseline YOLOv11. These gains were achieved with a smaller model size (5.08 MB) and a processing speed of 135.2 frames per second (f/s). The model also performed well on the proprietary dataset, demonstrating strong generalization.

Conclusion

The results indicate that the improved algorithm achieves higher accuracy, real-time performance, and better generalization in lung nodule detection, highlighting its potential for clinical application in the early lung cancer diagnosis.