Details of Research Outputs

Accurate detection and discovery of early lung cancer is the most effective measure to reduce lung cancer mortality with high clinical value. However, existing common object detectors show unsatisfactory detection accuracy for pulmonary nodule detection, due to the textureless appearance and small size of nodules. To address the textureless appearance problem, we propose a dedicated Nodule-Learning C3 module, which helps to extract more informative structures from limited textures of nodules. Considering that nodules' sizes are small, we further design a tiny object detection layer that performs object detection on larger feature maps, where more nodule features are preserved. Moreover, the balance between speed and accuracy is also critical for the pulmonary nodule diagnostic system. Therefore, we choose the famous one-stage detection framework YOLO [13] as our baseline and implement our proposed module and layer based on it. Extensive experimental results on the widely used benchmark LUNA16 demonstrate the superior performance of our method, in terms of both accuracy and speed. Specifically, our model improves the mAP accuracy by over and is faster than the YOLO baseline. © 2023 ACM.