Analysis and measurement of Tuberculin skin test induration using deep neural network

Abstract

The World Health Organization (WHO) posited that tuberculosis (TB) is among the world’s ten greatest causes of mortality. Early case identification and timely treatment could minimize TB morbidity and death rates. This study adopts the UNets model for automatically detecting TB in subjects by using a deep neural network to assess the size of induration after tuberculin was injected into their hands. In order to do this, two neural network models were fine-tuned utilizing pre-learned weights from the 2012 ILSVRC ImageNet. Algorithms were developed to perform semantic segmentation of induration and compare it to that of a reference object of a known dimension. This was used to classify the status of the subject as either positive or negative. A series of experiments performed demonstrated that the optimal selection of neural network hyperparameters may provide a satisfactorily high F1 score of up to 0.977.