Análise da influência de compressão, escalonamento não uniforme e aumento de dados na classificação automática de micro-organismos em imagens biomédicas utilizando deep learning

Abstract

Correct identification and classification of microorganisms is an important task as it helps in detecting and preventing disease outbreaks, tracking antibiotic resistance, and monitoring disease trends to see if prevention measures are working. Usually, specialized professionals are responsible for the classification of microorganisms through visual examination of microscope images. Due to the high error rates, common in manual processes, especially in the most complex ones, one way to improve quality is through the automation of this task using new technologies. Studies that apply deep learning to classify microorganisms through convolutional neural networks (CNN) have been showing good results, however, no studies in the literature on this subject show the influence of factors such as image compression, non-uniform data scaling, and data augmentation, especially using small sets of microorganism images, in the performance of CNNs. In view of this, the objective of this research is to analyze the influence of such factors by carrying out experiments with the AlexNet and DenseNet-121 CNN architectures, which have already shown good results in other research on this topic. The results obtained in the conducted experiments showed that some factors, such as the type of data compression used and the non-uniform scaling applied to the images negatively affect the performance of the CNNs, while the increase of data using image processing techniques such as mirroring, rotation and noise injection improve the performance of CNNs. Nevertheless, the use of training sets created from this knowledge and the use of optimized configuration parameters of the CNNs allowed obtaining accuracies of 98.61% with AlexNet and 99.82% with DenseNet-121, which are higher than those reported in the state of the art. Thus, this research can help reduce dataset preparation time and RNC training, reducing the computational cost and increasing the consistency and accuracy of microorganism identification.