Whole Mammography Diagnosis via Multi-instance Supervised Discriminative Localization and Classification

Abstract

Precise mammography diagnosis plays a vital role in breast cancer management, especially in identifying malignancy with computer assistance. Due to high resolution, large image size, and small lesion region, it is challenging to localize lesions while classifying the whole mammography, which also renders difficulty for annotating mammography datasets and balancing tumor and normal background regions for training. To fully use local lesion information and macroscopic malignancy information, we propose a two-step mammography classification method based on multi-instance learning. In step one, a multi-task encoder-decoder architecture (mt-ConvNext-Unet) is employed for instance-level lesion localization and lesion type classification. To enhance the ability of feature extraction, we adopt ConvNext as the encoder, and added normalization layer and scSE attention blocks in the decoder to strengthen localization ability of small lesions. A classification branch is used after the encoder to jointly train lesion classification and segmentation. The instance-based outputs are merged into the image-level both for segmentation and classification (SegMap and ClsMap). In step two, a whole mammography classification model is applied for breast-level cancer diagnosis by combining the results of CC and MLO views with EfficientNet. Experimental results on the open dataset show that our method not only accurately classifies breast cancer on mammography but also highlights the suspicious regions.