<p>In recent years, large-kernel Convolutional Neural Networks (CNNs) have received significant attention in the classification of COVID-19 X-ray images. However, there are still unresolved issues. Existing large-kernel convolutional neural networks tend to overly emphasize local perceptual abilities, while the question of whether these networks possess strong general perceptual ability at multiple scales in the visual domain requires further exploration. Additionally, current neural network architecture designs primarily focus on spatial feature extraction from static images and have limited capability to understand temporal dynamic information in image sequences. To overcome these challenges, we introduce TMscNet, a tailored deep learning model for COVID-19 X-ray image classification. TMscNet introduces a new convolutional structure called DiverseConvBlock on top of the ConvNeXt network. It employs a novel self-tuning algorithm to fuse feature maps generated from convolutions with different kernel sizes. Furthermore, we employ Global Response Normalization to concurrently capture contextual and global image features, thus enhancing scale-based feature perception and extraction efficiency. Additionally, we introduce the Temporal Convolutional Network (TCN) to enable better capture and utilization of dynamic information in image sequences by leveraging its efficiency in handling long-term dependencies. Our model achieves high accuracy, specificity, and sensitivity on the publicly accessible Chest X-Ray Dataset and COVID-19 Radiography Database. In the classification of the first dataset, our model demonstrated an accuracy of 98.14%, a specificity of 98.79%, and a sensitivity of 98.29%. Upon validating with the second dataset, it achieved an accuracy of 98.05%, with specificity and sensitivity reaching 99.01% and 98.12%, respectively.These impressive results confirm the outstanding performance of our proposed network in COVID-19 and pneumonia image classification tasks.</p>