Mammogram classification based on a novel convolutional neural network with efficient channel attention. (November 2022)

• Record Type: Journal Article
• Authors: Lou, Qiong; Li, Yingying; Qian, Yaguan; Lu, Fang; Ma, Jinlian
• Is Part Of: Computers in biology and medicine. Volume 150(2022)
• Publication Date: 2022-11
• Subjects: Mammogram classification -- Efficient channel attention -- ResNet50 -- Focal loss -- Transfer learning; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2022.106082
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Early accurate mammography screening and diagnosis can reduce the mortality of breast cancer. Although CNN-based breast cancer computer-aided diagnosis (CAD) systems have achieved significant results in recent years, precise diagnosis of lesions in mammogram remains a challenge due to low signal-to-noise ratio (SNR) and physiological characteristics. Many researchers achieved excellent performance in detecting mammographic images by inputting region of interest (ROI) annotations while ROI annotations require a great quantity of manual labor, time and resources. We propose a two-stage method that combines images preprocessing and model optimization to address the aforementioned challenges. Firstly, we propose the breast database preprocess (BDP) method to preprocess INbreast then we get INbreast † . The only label we need is benign or malignant label of one mammogram, not manual labeling such as ROI annotations. Secondly, we apply focal loss to ECA-Net50 which is an improved model based on ResNet50 with efficient channel attention (ECA) module. Our method can adaptively extract the key features of mammograms, meanwhile solving the problem of hard-to-classify samples and unbalanced categories. The AUC value of our method on INbreast † is 0.960, accuracy is 0.929, Recall is 0.928. The precision of our method on INbreast † is 0.883 which improved by 0.254 compared to ResNet50. In addition, we use Grad-CAM to visualize the effect of our model. The visualized heatmaps extracted by our method can focus more on lesion regions. Both numerical and visualized experiments demonstrate that our method achieves satisfactory performance. Highlights: Aim to assist clinical breast cancer research. Propose a two-stage method that combines mammogram preprocessing and model optimization. Achieve satisfactory performance both in numerical and visualized experiments.