Deep Learning in Breast Cancer Imaging: A Decade of Progress and Future Directions

TL;DR

Breast cancer imaging presents a rich, multi-modal domain where deep learning has progressed across screening, diagnosis, treatment response, and prognosis over the last decade. The paper provides a structured synthesis of methods (classification, detection, segmentation) and DL paradigms (supervised, weakly supervised, SSL, transfer, multimodal), organized by imaging modality (mammography, ultrasound, MRI, pathology). It highlights key trends such as multi-view and multi-parametric fusion, MIL and CAM-based weak supervision, and a shift toward prognostic and biomarker-driven AI, while emphasizing data scarcity, the need for external validation, and regulatory considerations. The work underscores the importance of robust, explainable, and personalized AI, and points to federated learning, synthetic data, and open multimodal datasets as critical avenues for future impact.

Abstract

Breast cancer has reached the highest incidence rate worldwide among all malignancies since 2020. Breast imaging plays a significant role in early diagnosis and intervention to improve the outcome of breast cancer patients. In the past decade, deep learning has shown remarkable progress in breast cancer imaging analysis, holding great promise in interpreting the rich information and complex context of breast imaging modalities. Considering the rapid improvement in deep learning technology and the increasing severity of breast cancer, it is critical to summarize past progress and identify future challenges to be addressed. This paper provides an extensive review of deep learning-based breast cancer imaging research, covering studies on mammogram, ultrasound, magnetic resonance imaging, and digital pathology images over the past decade. The major deep learning methods and applications on imaging-based screening, diagnosis, treatment response prediction, and prognosis are elaborated and discussed. Drawn from the findings of this survey, we present a comprehensive discussion of the challenges and potential avenues for future research in deep learning-based breast cancer imaging.

Figures (3)

• Figure 1: Overview of deep learning in breast cancer imaging. Typical imaging techniques include mammogram, ultrasound, magnetic resonance imaging (MRI), and pathology images. Deep learning is often used for screening, diagnosis, treatment response prediction, and prognosis.
• Figure 2: Number of representative papers on deep learning for breast cancer imaging published from 2012 to 2022.
• Figure 3: Brief illustration of deep learning models, taking mammogram as an example. (a) A typical classification network that uses convolutional and pooling to downsample the image while expanding the channels of features. The final feature maps will be pooled into a feature vector, and often a fully-connected layer can be used to conduct the classification based on the feature vector. Typical feature maps extracted by a ResNet-18 pre-trained on ImageNet from layers 1, 7, and 17 are shown in (b), (c), and (d), respectively. (e) A typical detection network. The downsampling workflow often follows the classification network. Then, the feature maps are upsampled, the multi-scale features are fed into a region proposal network (RPN) for region proposal generation, and a region-wise classification is performed to determine the final output. (f) A typical segmentation network. The downsampling workflow could follow the classification network. Then, the feature maps are upsampled several times and concatenated with the shallow-layer features. The final results are obtained based on pixel-wise classification on the largest feature map. All the models are optimized with backpropagation rumelhart1986learning.