Lung Nodule Detection Based on Spike-Driven Self-Attention YOLO

Spiking neural networks (SNNs) are brain-inspired energyefficient models. Recent researches have shown great success in achieving satisfying results on object detection based on SNNs. However, most researches only aim to detect objects in natural images. Few researches concentrate on medical applications. Lung nodule detection is one of meaningful tasks in medical image analysis, which plays an important role in the diagnosis of lung cancer. You Only Look Once (YOLO) is less efficient in detecting small objects such as nodules. To address this problem, we design a Spike-Driven Self-Attention YOLO (SDSA-YOLO) model for nodule detection by integrating spike-driven self-attention (SDSA) with the YOLO model. SDSA can provide more positional information of nodules by giving more weight to features of nodules according to correlation of features and helps detection heads find nodules. Our model is directly-trained and fully spike-driven, which can avoid negative impacts caused by ANN-SNN conversion and non-spike calculation. Experiments are carried out on the LUNA16 dataset, and the results show that our method achieves better performance than the baseline models.