Lung cancer remains one of the malignant tumors with the highest global morbidity and mortality rates. Detecting pulmonary nodules in computed tomography (CT) images is essential for early lung cancer screening. However, traditional detection methods often suffer from low accuracy and efficiency, limiting their clinical effectiveness. This study aims to devise an advanced deep-learning framework capable of achieving high-precision, rapid identification of pulmonary nodules in CT imaging, thereby facilitating earlier and more accurate diagnosis of lung cancer. To address these issues, this paper proposes an improved deep-learning framework named YOLOv8-BCD, based on YOLOv8 and integrating the BiFormer attention mechanism, Content-Aware ReAssembly of Features (CARAFE) up-sampling method, and Depth-wise Over-Parameterized Depth-wise Convolution (DO-DConv) enhanced convolution. To overcome common challenges such as low resolution, noise, and artifacts in lung CT images, the model employs Super-Resolution Generative Adversarial Network (SRGAN)-based image enhancement during preprocessing. The BiFormer attention mechanism is introduced into the backbone to enhance feature extraction capabilities, particularly for small nodules, while CARAFE and DO-DConv modules are incorporated into the head to optimize feature fusion efficiency and reduce computational complexity. Experimental comparisons using 550 CT images from the LUng Nodule Analysis 2016 dataset (LUNA16 dataset) demonstrated that the proposed YOLOv8-BCD achieved detection accuracy and mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5 (mAP0.5) of 86.4% and 88.3%, respectively, surpassing YOLOv8 by 2.2% in accuracy, 4.5% in mAP0.5. Additional evaluation on the external TianChi lung nodule dataset further confirmed the model's generalization capability, achieving an mAP0.5 of 83.8% and mAP0.5-0.95 of 43.9% with an inference speed of 98 frames per second (FPS). The YOLOv8-BCD model effectively assists clinicians by significantly reducing interpretation time, improving diagnostic accuracy, and minimizing the risk of missed diagnoses, thereby enhancing patient outcomes.
Zhu et al. (Thu,) studied this question.