Aprimoramento da detecção de câncer em bases de imagens limitadas: uma abordagem baseada em ensemble e aumento de dados

Abstract

Recently, Deep Learning models have gained a lot of prominence in the area of image analysis, but they need a lot of data to perform well. This makes the application in medical databases difficult, since there is a difficulty in obtaining many observations equally, between different cases, such as pathological and normal. For small databases, models of learning transfer, ensemble and data augmentation can be used to have better performance in the task of classifying images, while for unbalanced databases, resampling techniques such as undersampling and oversampling can be used. In this work we propose a new approach based on the ensemble of transfer learning models with different weights to improve the prediction of minority class observations in small unbalanced databases considering resampling and data augmentation. At the end, two experiments are carried out, the first with the database from the Hospital A.C.Camargo Cancer Center, which is an oncology hospital specialized in the diagnosis, treatment and research of cancer. This database has never been used for any type of analysis and its objective is to classify mammography images as malignant or benign. The second experiment is a database of benign or malignant skin cancer images from ISIC 2016. From the results obtained, it is noted that the combinations with the best metrics are obtained with the use of resampling and the worst without the use of this technique. This proves that the use of the resampling methodology (undersampling and oversampling) improves the performance of the combination in predicting the minority class. Yet, with the use of data augmentation, the proposed combinations have a better performance than not using this technique.