A LIGHT WEIGHT MULTI HEAD SELF ATTENTION MECHANISM TO CONVOLUTION NEURAL NETWORK TOWARDS DOMAIN INCREMENTAL CHEST X-RAY IMAGE CLASSIFICATION

In recent year, Chest X-ray images have been widely adopted towards pathological diagnosis of the respiratory diseases and pulmonary diseases. Advances in computer vision approaches and artificial intelligence offers promising solutions with high accuracy and low computation overhead to disease classification and severity estimation. Traditionally Chest X-ray images classification has been performed using machine learning and deep learning model architectures to provide excellent performances as it incorporates domain incremental learning. Despite of several advantages of those models, domain incremental learning faces catastrophic forgetting challenges due to large domain gaps. In this paper, light weight multihead self-attention mechanism which is considered as transformer architecture is incorporated into ResNet 50 based convolution neural network to fine tune strategies to learn and calculate the domain invariant features and domain specific features. Especially attention mechanism performs feature fusion among different domains to prevent forgetting. Further multi-label contrastive loss is built and trained as loss function to enhance class distinctiveness of the disease classes. Initially Convolution layer of the ResNet architecture is applied to extract the hierarchical feature from X-ray images. Next, Extracted features were projected to pooling layer incorporating the attention mechanism. Attention mechanism use attention coefficient to compute the domain invariant features and domain specific features to establish feature maps. Those feature map of different domains are aggregated and classified in the fully connected layer composed of activation function, normalization function, softmax function and loss function. Activation function using ReLu function aggregates the feature maps while softmax function using Naïve bayes classifier classify the feature map into disease categories as COVID-19 variants, Normal, Viral Pneumonia, and Lung Opacity. Experimental results and performance results substantiate the effectiveness of the current architecture with accuracy of 98.9% and surpassing the baseline methods exclusively in classifying the Chest x –ray images into multiple classes. Finally results emphasize the significance of the transformer architecture for optimizing the conventional deep learning architecture in the medical imaging.