Efficient Esophageal Lesion Detection using Polarization Regularized Network Slimming

Early detection of esophageal cancer is of great significance for improving the survival rate of esophageal cancer patients, and the esophageal cancer detection model based on deep learning can be used as a powerful tool to improve the detection rate. YOLO-V5 has been one of the leading methods for esophageal lesion detection at present. However, these networks cannot be effectively adopted for real-time applications as they are parameter-heavy, computationally complex and difficult to meet the efficiency requirements of clinical scenarios. To this end, we propose an end-to-end esophageal cancer lesion detection network based on polarization regularization and network slimming, which is named EPS YOLO-V5. First, we build a feature extraction network by utilizing modules from GhostNet. Then, we optimize the scale factors in BN layers with sparsity and polarization regularizations during network training, which is beneficial to determining the pruning threshold by widening the distance gap between the pruned neurons and the preserved neurons, thus enables automatic search of the pruning threshold. Lastly, the pruned network is fine-tuned to regain the performance. We test the proposed EPS YOLO-V5 on a collected esophageal image dataset, and the experimental results show that the proposed method reduces the number of parameters by 5 X, decrease the size of model by 8x, and improve the inference speed by 4.5 X while obtaining comparable performance compared with the original YOLO-V5 for esophageal lesion detection.