CED-Net: A Generalized Deep Wide Model for Covid Detection

Attention-based deep learning models in medical imaging have demonstrated appreciable results, with wide models focusing on memorization and deep models on generalization. COVID detection involves different types of CT scan and X-ray images of varying severity, demands a generalized model with high sensitivity. However, the gener- alization of deep learning models with selective specialization for specific data is still challenging. We introduce a deep-wide structured block archi- tecture (called CED-Net), involving a convolution-attention mechanism that combines the strengths of both wide linear and deep neural network models. The dimension scaling of the proposed model by EfficientNet with pre-trained knowledge of DenseNet complements the multi-headed self-attention of CNNs meet Vision Transformers model to effectively generalize the model for varying lesions on X-ray and CT-scan images. By leveraging feature retention and impedance avoidance, our proposed model has demonstrated good performance on COVID datasets.