This text was initially printed right here
J Med Imaging (Bellingham). 2021 Jan; 8 (Complement 1): 017503. doi: 10.1117 / 1.JMI.8.S1.017503. Epub 2021 August 23
ABSTRACT
Goal: The goal of this examine is to develop and consider a totally automated, deep learning-based methodology for detecting a COVID-19 an infection utilizing chest x-rays. Strategy: The proposed mannequin was developed by changing the final classifier layer in DenseNet201 with a brand new community consisting of a world averaging layer, a stack normalization layer, a dense layer with ReLU activation and a last classification layer. Then we did an end-to-end exercise with the initially pre-trained weights on all shifts. Our mannequin was skilled with a complete of 8644 photos, with 4000 photos every in regular and pneumonia instances and 644 in COVID-19 instances representing a big actual information set. The proposed methodology was assessed primarily based on accuracy, sensitivity, specificity, ROC curve and F 1 rating utilizing a take a look at information set with 1729 photos (129 COVID-19, 800 regular and 800 pneumonia). As a benchmark, we additionally in contrast the outcomes of our methodology with these of seven state-of-the-art pre-trained fashions and a light-weight CNN structure developed from scratch. Outcomes: The proposed mannequin primarily based on DenseNet201 was capable of obtain an accuracy of 94% within the detection of COVID-19 and an total accuracy of 92.19%. The mannequin was capable of obtain an AUC of 0.99 for COVID-19, 0.97 for regular, and 0.97 for pneumonia. The mannequin outperformed different fashions by way of total accuracy, sensitivity, and specificity. Conclusions: Our proposed automated diagnostic mannequin confirmed an accuracy of 94% when initially screening COVID-19 sufferers and an total accuracy of 92.19% when utilizing chest x-rays.
PMID: 34435075 | PMC: PMC8382139 | DOI: 10.1117 / 1.JMI.8.S1.017503
