Journal
IEEE ACCESS
Volume 9, Issue -, Pages 905-922Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/ACCESS.2020.3044625
Keywords
Atrous convolution; faster-RCNN; histopathology; multiscale learning; mitosis detection; wavelet transform
Categories
Funding
- Shenzhen Fundamental Research Project [JCYJ20170412151226061, JCYJ20170808110410773, JCYJ20180507182241622]
Ask authors/readers for more resources
The article introduces a novel framework called SmallMitosis to detect very small mitotic cells in breast cancer tumors, using an A-FCN model and MS-RCNN detector. By combining atrous convolution and multiscale techniques, the proposed scheme outperforms state-of-the-art approaches in detecting small size mitosis from weakly labeled datasets.
Mitotic figure count acts as a proliferative marker to measure aggressiveness of the breast cancer tumor. In this article, we have proposed a novel framework named SmallMitosis to detect mitotic cells particularly very small size mitosis from hematoxylin and eosin (H&E) stained breast histology images. SmallMitosis framework consists of an atrous fully convolution based segmentation (A-FCN) model and a deep multiscale (MS-RCNN) detector. In intended A-FCN model, the concept of atrous convolution helps to estimate mitosis mask and bounding box annotations of very small size mitotic cells. Meanwhile, the architecture of MS-RCNN internally lifts poor representations of small mitosis to super-resolved ones, that are similar to real large mitosis thus more discriminative for detection of small size blurred mitotic cells, as well as a fully convolution layer at detection stage, decreases computational cost. The A-FCN model trained on fully labeled mitosis datasets (all pixels of mitosis are labeled) is applied on weakly labeled datasets (only centroid pixel is labeled) to obtain mitosis mask and bounding box annotations. Using these estimated bounding box annotations, MS-RCNN detector is trained to detect small size mitosis from weakly labeled datasets. The performance of the proposed scheme is tested on three publicly available mitosis datasets, namely ICPR 2012, ICPR 2014, and AMIDA13. On challenging ICPR 2012 dataset, we obtained F score of 0.902, outperforming all prior detection systems significantly. On ICPR 2014 and AMIDA13 datasets, we achieved a 0.495 and 0.644 F score respectively. The results demonstrated that our method impressively outperforms state-of-the-art approaches. SmallMitosis is available at https://github.com/tasleem-hello/SmallMitosis.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available