Content-Based Retrieval of COVID-19 Affected Chest X-rays with Siamese CNN

Abstract

AbstractIn December 2019, the first reported case of COVID-19 was brought to notice in Wuhan, China. The virus has novel characteristics, its harshness is unpredictable, its transmission ability is extremely powerful, and its incubation period is comparatively larger. Thus the outbreak emerged as a pandemic worldwide. World health and socio-economy is getting continually affected by COVID-19 since its outbreak. It will be easier to handle the situation if an automated diagnostic system is developed, capable of separating COVID-19 affected images from bulk images obtained from a mass screening process. Kaggle’s online chest X-Ray image dataset has been considered for this work evaluation. Healthy and COVID-19 affected chest X-Ray images were used for evaluating the performance of content-based image retrieval. Image retrieval has been carried out based on the absolute difference between the encoded features of twin images obtained from the Siamese Convolutional Neural Network (SCNN). The retrieval performance is awe-inspiring as the Siamese network used for retrieval is a relatively shallow network. SCNN does not require resource-hungry training with huge samples as part of its underlying implementation characteristics. The execution time is also very encouraging as the simplicity of the method is concerned. The method achieves 94% average precision and 100% average reciprocal rank while rank = 5 has been considered. Till now, no work has been reported on content-based retrieval of COVID-19 chest X-Ray images. Thus, a comparative study of evaluation metrics and execution time requirements of similar work could not be provided.KeywordsChest X-ray imagesCOVID-19Siamese convolutional neural network (SCNN)Siamese distance measurementMean average precision (MAP)Mean reciprocal rank (MRR)