Mitosis Detection from Breast Histopathology Images Using Mask RCNN

Histopathology images are commonly used for cancer detection and prognosis as these images provide tissue level information which allow a pathologist to view the tumor cells and quantify them. This information can be used to identify any proliferative activity as well as determine the aggressiveness of the disease based on the number of mitotic or tumor cells as well as the ongoing stage of mitosis. This can be used for early diagnosis of breast cancer and can contribute towards reducing the mortality rates. However, manually identifying tumor cells is labor intensive and is prone to challenges like class imbalance, variability, etc. To overcome these challenges and reduce the dependency on pathologists, this work leveraged the benefits of deep learning techniques and developed a framework for mitosis detection from breast histopathology images using Mask Region-based Convolutional Neural Network (Mask RCNN) model. An ROI based patch extraction technique was used to train and evaluate the model on GZMH dataset. The proposed framework achieved a precision of 67.36
^%
with recall of 63.33%, mAP of 0.589 and AUC-PR of 0.67, thereby outperforming other state-of-art models on the GZMH dataset.