Diagnosis of atrial fibrillation based on lightweight detail-semantic network

•A lightweight detail-semantic network is proposed in this paper. Using one-dimensional ECG as input, we construct detail path using depthwise convolution and point convolution, and extract deep information of AF by increasing the depth of the network.•The semantic path built by the hourglass residual module is used to compensate for the lack of information interaction in depthwise separable convolution and to ensure that the information learned by the network contains temporal order.•The hourglass residual module uses compressor theory to reduce the computation of the residual module while ensuring feature reuse.•In order to achieve mutual assistance between detail path and semantic path, this paper uses an improved cross-guidance mechanism to guide the information transfer and information fusion between different paths.•The network has only 0.2 M parametric count and 327.36 M computational volume. It achieves 99.57% accuracy on the MIT-BIH Atrial Fibrillation Database and 90.89% accuracy on the clinical dataset.