Multi-Scale Deep Neural Network for Mitosis Detection in Histological Images

Mitotic figure detection in breast cancer images plays an important role to measure aggressiveness of the cancer tumor. Currently, in clinic environment the pathologist visualized the multiple high power fields (HPFs) on a glass slide under super microscope which is an extremely tedious and time consuming process. Development of the automatic mitotic detection methods is need of time, however it also bears, scale invariance, deficiency of data, improper image staining and sample class unbalanced dilemma. These limitations are however; prohibit the automatic histopathology image analysis to be applied in clinical practice. In this paper, an automatic domain agnostic deep multi-scale fused fully convolutional neural network (MFF-CNN) is presented to detect mitoses in Hematoxylin and eosin (H&E) images. The intended model fuses the multi-level and multi-scale features and context information for accurate mitotic count and in training phase multi-step fine-tuning strategy is used to reduce the over-fitting. Moreover, the training image samples efficiently built by stain normalized the poorly stained (H&E) images and by applying an automatic sample selection strategy. Preliminarily validation on the public MITOS-ATYPIA-14 challenge dataset, demonstrate the efficiency of proposed work. The proposed method achieves better performance in term of detection accuracy with an acceptable detection speed compared to other state-of-the-art designs.