Abstract:
Breast arterial calcifications (BAC) are a type of calcification commonly observed on mammograms and are generally considered benign and not associated with breast cancer. However, there is accumulating observational evidence of an association between BAC and cardiovascular disease, the leading cause of death in women. We present a deep learning method that could assist radiologists in detecting and quantifying BAC in synthesized 2D mammograms. We present a recurrent attention U-Net model consisting of encoder and decoder modules that include multiple blocks that each use a recurrent mechanism, a recurrent mechanism, and an attention module between them. The model also includes a skip connection between the encoder and the decoder, similar to a U-shaped network. The attention module was used to enhance the capture of long-range dependencies and enable the network to effectively classify BAC from the background, whereas the recurrent blocks ensured better feature representation. The model was evaluated using a dataset containing 2,000 synthesized 2D mammogram images. We obtained 99.8861% overall accuracy, 69.6107% sensitivity, 66.5758% F-1 score, and 59.5498% Jaccard coefficient, respectively. The presented model achieved promising performance compared with related models.
摘要:
乳腺动脉钙化(BAC)是一种通常在乳房X线照片上观察到的钙化,通常被认为是良性的,与乳腺癌无关。然而,越来越多的观察证据表明BAC与心血管疾病之间存在关联,女性死亡的主要原因。我们提出了一种深度学习方法,可以帮助放射科医生在合成的2D乳房X线照片中检测和量化BAC。我们提出了一个循环注意力U-Net模型,该模型由编码器和解码器模块组成,其中包括多个块,每个块都使用循环机制,一种反复出现的机制,和他们之间的注意模块。该模型还包括编码器和解码器之间的跳过连接,类似于U形网络。注意模块用于增强对远程依赖关系的捕获,并使网络能够从背景中有效地对BAC进行分类,而循环块确保了更好的特征表示。使用包含2,000个合成的2D乳房X线照片图像的数据集评估模型。我们获得了99.8861%的总体准确率,69.6107%灵敏度,66.5758%F-1得分,和59.5498%的Jaccard系数,分别。与相关模型相比,提出的模型取得了有希望的性能。