Automatic detection metastasis in breast histopathological images based on ensemble learning and color adjustment

Breast cancer is a common neoplasm among women. The cure of the disease depends on early identification and treatment of the tumor to avoid advanced stages such as metastasis of the initial lesion. Recently, studies have shown that computational methods could help specialists in this process and even provide advantages over the traditional analysis method. Thus, this work proposes a tumor cell detection method based on an ensemble of convolutional neural networks (CNN) trained with normalized images using different color adjustment techniques. Analyses of eight different color spaces and their channels, and two color normalization methods were performed to reduce the effects caused by color variation of histopathological images on the generalization of predictive models. The proposed approach evaluated six CNN architectures with color adjustment methods to define which ones improve and preserve important characteristics for the classification task. The ensemble that constitutes the proposed method comprises three models of the VGG-19 architecture trained in images generated through from the color space HSV, from the color channel RED of the RGB images, and RGB images normalized with Reinhard method, respectively. This approach was evaluated using a public image database containing 327,680 histopathological images extracted from breast tissue. The method achieved promising results, with an accuracy of 0.9193 and an AUC of 0.9772. These results demonstrate that a combination of color adjustment techniques produces better results than applying the techniques individually.