Multiscale Context-Cascaded Ensemble Framework (MsC<sup>2</sup>EF): Application to Breast Histopathological Image

Zhenyuan Ning, Xinsen Zhang, Chao Tu, Qianjin Feng, Yu Zhang

IEEE Access · 2019 · 10 citations · 42 references

DOIFull text

Open access

Concepts

Abstract

Microscopic analysis of breast cell and tissue is a critical step in the definitive diagnosis of breast cancer. However, it's time-consuming and fatigable for histopathologists to find the diagnostic characteristic of cell and tissue in breast histopathological image through multiple magnification scannings. Many computer-aided studies, including traditional machine learning and deep learning approaches, have been conducted to efficiently assist histopathologists in making diagnostic decision. However, precision and complexity of such approaches remain challenging. In this work, we propose and evaluate a new framework, called multiscale context-cascaded ensemble framework (MsC <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> EF), to classify breast histopathological images. The model based on MsC <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> EF exhibits a higher precision than traditional machine learning. Meantime, it is more efficient and hardware-independent compared with deep learning approaches. The MsC <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> EF consists of the input, cascade, and decision layers. The input layer comprises a feature extractor and a spatial pyramid of image to execute feature input from coarse to fine scales. Four ensemble channels are stacked in a parallel manner as the cascade layer to select and transfer contextual feature iteratively and adaptively. For the decision layer, kernel fusion-based method is integrated to perform classification of breast histopathological image by fusing four different feature spaces. Our proposed method has been evaluated on an open dataset. The experimental result shows that MsC <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> EF obtains a good classification performance (Accuracy at patch level: 0.948±0.016; accuracy at patient level: 0.981±0.016), indicating its potential application to the classification of breast histopathologist images.

References

42