Strongly Supervised Mitosis Detection In Breast Histopathology Images Using Weak Labels

Abstract

Mitosis detection plays an import role in tumor grading of breast cancer. Automatic mitosis detection liberates pathologists from the time-consuming manual counting work. However, mitosis detection datasets are often provided only with centroid labels, limiting the performance of most deep learning methods. In this paper, we propose a novel method for mitosis detection to address this problem. First, we generate pixel-level labels directly from origin centroid labels with a gradient changing threshold approach. Then we apply a ResNet-based FCN to detect mitotic nuclei. Low confidence and tiny areas are removed from the prediction map to produce the final detections. By transforming the weak labels into strong labels, our method achieves $\\mathrm{F}_{1} -$scores of 0.692, 0.608 and 0.805 on three public datasets, AMIDA 2013, ICPR 2014 and TUPAC 2016, respectively, outperforming all other state-of-the-art methods and showing great potential for clinical deployment.