YOLO-ED: An efficient lung cancer detection model based on improved YOLOv8

In recent years, You Only Look Once (YOLO) models have gradually been applied to medical image object detection tasks due to their good scalability and excellent generalization performance, bringing new perspectives and approaches to this field. However, existing models overlook the impact of numerous consecutive convolutions and the sampling blur caused by bilinear interpolation, resulting in excessive computational costs and insufficient precision in object detection. To address these problems, we propose a YOLOv8-based model using Efficient modulation and dynamic upsampling (YOLO-ED) to detect lung cancer in CT images. Specifically, we incorporate two innovative modules, the Efficient Modulation module and the DySample module, into the YOLOv8 model. The Efficient Modulation module employs a weighted fusion strategy to extract features from input CT images, effectively reducing model parameters and computational overhead. Furthermore, the DySample module is designed to replace the conventional upsampling component in YOLO, thereby mitigating information loss when expanding feature maps. The dynamic bilinear interpolation introduced by this module increases random bias, which helps minimize errors in feature extraction. To validate the effectiveness of YOLO-ED, we compared it with baselines on the LUNG-PET-CT-DX lung cancer diagnosis dataset and the LUNA16 lung nodule dataset. The results show that YOLO-ED significantly improves precision and reduces computational cost on these two datasets, demonstrating its superiority in the detection of medical images.