KansNet: Kolmogorov–Arnold Networks and multi slice partition channel priority attention in convolutional neural network for lung nodule detection

Globally, lung cancer ranks as the primary reason for fatalities associated with cancer. Accurate detection of pulmonary nodules in computed tomography (CT) images is crucial for the early diagnosis of lung cancer, which significantly improves patient survival rates. Nonetheless, for seasoned radiologists, identifying these nodules among the extensive collection of CT images continues to be a difficult endeavor. In this study, we propose a framework called KansNet for the detection of pulmonary nodules. We have designed a partial attention module based on KAN (Kolmogorov–Arnold Networks) integrating global representation learning capabilities into the CNN and also reveals the excellent potential of KAN in the field of CT lung nodule detection. Furthermore, we present an innovative adaptive feature fusion module, paired with a multi slice partition channel priority attention module, which improves the network’s capacity to manage small features and identify small nodules. In order to thoroughly assess the effectiveness of different models for detecting pulmonary nodules, we carried out experiments utilizing the publicly accessible LUNA16 dataset. Our approach surpasses alternative detection algorithms in terms of CPM, achieving a CPM score of 90.32%, which represents an enhancement of 2.11%. Importantly, our approach demonstrated greater sensitivity at low false positive rates. Additionally, we achieved a 2.21% improvement in the AP metric, reaching an AP score of 89.19%. Moreover, our model also achieved excellent results on the TIANCHI dataset, with CPM and AP values reaching 69.42% and 67.99%, respectively, representing improvements of 1.51% and 2.23%. •A partial attention-based KAN module is designed, and a new model for CT lung nodule detection is proposed.•A 3D adaptive feature fusion module is introduced, and a channel priority attention module is designed.•The proposed model outperforms previous CNN-based approaches on public datasets.