Multiple Visual Fields Cascaded Convolutional Neural Network For Breast Cancer Detection

This paper proposes a novel multiple visual fields cascaded convolutional neural network MVF-CasCNN to detect breast cancer metastases in the gigapixel whole slide image WSI. Here visual field is the total area of what the classifier can perceive from a particular input. Firstly, we perform patch-level classification using a large visual field CNN to coarsely locate possible lesion regions in a WSI. Then, a small visual field CNN is adopted to finely identify these tumor candidate areas and generate the final lesion probability heatmap for the WSI. Compared with single visual field based models, MVF-CasCNN achieves much higher performance in both slide classification and lesion localization. We also present a tight definition of tumor patch and an efficient relatively hard example mining method to enhance our network. Experimental results show that our method can surpass pathologist's level and achieve the state-of-the-art performance on public dataset Camelyon16.