Melhorando o desempenho na classificação de imagens médicas: uma análise comparativa de técnicas de aumento de dados

Abstract

Classification and segmentation methods for medical image analysis based on convolutional neural networks (CNNs) have achieved state-of-the-art accuracy. However, the scarcity of labeled medical data available for supervised training of CNNs can compromise their performance. To overcome this limitation, data augmentation methods have been successfully employed. These methods involve applying transformations to the input data, such as rotation, translation, scaling, and mirroring, to generate synthetic images that enrich the training set. Additionally, they can be applied to different types of data, including applications involving various types of medical images. The objective of this study is to evaluate the effectiveness of four automatic data augmentation methods: Baseline, AugMix, RandAugment, and TrivialAugment. We aim to assess the performance of a CNN model in classifying X-ray images of COVID-19 pneumonia, utilizing a dataset divided into four distinct classes. The dataset will be augmented using four different data augmentation scenarios. Subsequently, the results obtained from the training and validation processes will be compared and analyzed. The results showed that the AugMix method achieved the best performance among the evaluated methods, with an F1-Score of 0.910368, approximately 5% higher than the baseline method. The RandAugment and TrivialAugment methods followed, with F1-Scores of 0.896375 and 0.893653, respectively. Furthermore, it was noticed that compared to the baseline method, the evaluated methods were able to reduce model overfitting, improving its generalization capability for new cases. It is expected that the results of this study will contribute to a better understanding of the importance of data augmentation, assist in selecting the best data augmentation method for the domain of COVID-19 pneumonia X-ray images, and analyze the impact of different automatic data augmentation methods on the performance of supervised classification models for COVID-19 pneumonia image analysis.