Towards Multiple Enhancement Styles Generation in Mammography

Shilin Chen, Ji He, Jianhua Ma

Keywords: mammogram enhancement, deep learning

TL;DR: we present a deep learning (DL) framework to achieve multiple enhancement styles generation for mammogram enhancement.

Abstract:

Mammography is a well-established imaging modality for early detection and diagnosis of breast
cancer. The raw detector-obtained mammograms are difficult for radiologists to diagnose due to the
similarity between normal tissues and potential lesions in the attenuation level and thus mammogram
enhancement (ME) is significantly necessary. However, the enhanced mammograms obtained
with different mammography devices can be diverse in visualization due to different enhancement
algorithms adopted in these mammography devices. Different styles of enhanced mammograms
can provide different information of breast tissue and lesion, which might help radiologists to
screen breast cancer better. In this paper, we present a deep learning (DL) framework to achieve
multiple enhancement styles generation for mammogram enhancement. The presented DL framework
is denoted as DL-ME for simplicity. Specifically, the presented DL-ME is implemented with
a multi-scale cascaded residual convolutional neural network (MSC-ResNet), in which the output
in the coarser scale is used as a part of inputs in the finer scale to achieve optimal ME performance.
In addition, a switch map is input into the DL-ME model to control the enhancement style of the
outputs. To reveal the multiple enhancement styles generation ability of DL-ME for mammograms,
clinical mammographic data from mammography devices of three different manufacturers are used
in the work. The results show that the quality of the mammograms generated by our framework can
reach the level of clinical diagnosis and enhanced mammograms with different styles can provide
more information, which can help radiologists to efficiently screen breast cancers.

Track: short paper

Paper Type: methodological development