LNDCNN: A lung nodule detection model based on improved YOLOv7

The automatic detection of lung nodules is crucial for cancer screening. However, lung nodules typically occupy small areas in CT images with indistinct features, posing significant detection challenges. To address this, YOLOv7 is optimized and a lung nodule detection model is proposed. Firstly, considering the problem of feature loss caused by the PAFPN (Path Aggregation Feature Pyramid Network), an Enhanced Feature Pyramid Network (E_FPN) is introduced, which employs dilated convolutions in series to adjust the receptive field size and supplement contextual information. The dilation rate is designed to fully utilize fine-grained features. Then, an adaptive fusion strategy is proposed, enabling each detection head to adaptively integrate information from other detection heads, thereby achieving complementary utilization of detection head information. Finally, considering that the IoU (Intersection over Union) metric is very sensitive to the position change of small objects, the Gaussian Wasserstein distance metric is adopted to optimize the loss function and enhance target localization accuracy. Experiments are primarily conducted on the LUNA16 dataset. The optimized model, LNDCNN (Lung Nodule Detection Convolutional Neural Network), achieves an improvement of 11.42% in mAP₅₀, 6.17% in mAP${}_{50:90}$, and a 14.4% increase in Recall compared to YOLOv7. At the same time, we validate the algorithm’s generalization ability on the ALIBABA TianChi dataset. LNDCNN demonstrates consistent improvements with 5.68% higher mAP₅₀, 2.92% higher mAP${}_{50:90}$, and 8.61% higher Recall compared to YOLOv7, demonstrating promising generalization capability. The implementation of LNDCNN is publicly available at https://github.com/zlj-zlj/LNDCNN.