<![CDATA[Respiratory abnormality detection in chest X-ray images can be assisted and automated via machine learning models. Many of those models are very effective in detecting abnormalities, but they require large computational and memory resources. Running these models on edge devices often leads to delays in inference or crashes, as most of the devices have memory constraints. Therefore, a more edge-aware image classification model called Med-ConvNeXT is developed by keeping both memory and performance in mind. The proposed model introduces a kernel reparameterization technique called receptive kernel fusion, which combines useful information from kernels ( ) to calculate a receptive kernel. The kernel calculated by the above method can extract both lower- and higher-level features without increasing FLOPs or the network's parameter count. To further reduce the number of parameters and FLOPs, a novel channel pruning method, called Memory-based Channel Pruning, is introduced. This method compares the classification performance of different kernels and ranks them to eliminate the least important kernels. This reduced the total channel size from 96 to 64 in the first stage. Combining the proposed kernel reparameterization and channel pruning technique reduced the parameter count of the ConvNeXT model by up to 54.7% while achieving a classification accuracy of 96.53% for X-ray images. The proposed model achieved 93.7% sensitivity and 95.9% specificity with a parameter count of 12.96 million. The proposed model was implemented on Raspberry Pi 5 hardware and got an inference latency of 985ms using a 118 MB memory footprint. This performance is achieved with an input image dimension of a 512×512 X-ray image. The proposed model balances both image classification performance and edge implementation efficiency. The proposed model can achieve better performance on edge devices without the need for expensive server and cloud infrastructure]]>
