TLDR: MammoDINO is a new self-supervised learning framework for mammography, trained on 1.4 million images. It uses breast tissue-aware data augmentation and a cross-slice contrastive learning objective for 3D DBT to focus on clinically relevant features. This approach achieves state-of-the-art performance in breast cancer screening tasks across multiple datasets, offering an annotation-free foundation for improved diagnostic tools.
This paper introduces MammoDINO, a new self-supervised learning (SSL) framework specifically designed for mammographic images. While SSL has been very successful in general image processing, its application in medical imaging, especially mammography, has faced challenges due to limited data and unique characteristics of medical images. MammoDINO aims to overcome these hurdles by incorporating anatomical awareness into its training process.
Breast cancer is a major health concern for women, and mammograms are crucial for early detection. However, interpreting these images can be difficult even for experienced radiologists, leading to a need for better computer-aided diagnosis (CAD) tools. Traditional CAD tools often rely on supervised learning, which requires a lot of annotated data – something that is scarce in medical imaging. This is where foundation models and self-supervised learning come in, allowing models to learn from large amounts of unlabeled data.
Existing SSL methods, like DINOv2, often use random cropping and focus on individual 2D images. For mammograms, this can be problematic because much of the image might be non-informative background, and important clinical signals are often localized within breast tissue. Furthermore, 3D Digital Breast Tomosynthesis (DBT) involves multiple 2D slices where anatomical structures change smoothly, and standard 2D-focused SSL doesn’t fully capture this 3D context.
MammoDINO addresses these issues with two main innovations. First, it uses a “breast tissue aware augmentation sampler.” This means that during training, the model is specifically guided to focus on regions that contain breast tissue, ensuring it learns from clinically relevant areas and avoids wasting effort on irrelevant background. Second, it introduces a “3D DBT adjacent slice loss.” This novel approach helps the model understand the continuity of anatomical structures across adjacent 2D slices within a 3D DBT volume. By enforcing consistency between these nearby slices, MammoDINO learns more robust and meaningful representations of breast anatomy.
The MammoDINO framework was pretrained on a massive dataset of 1.4 million mammographic images, including both 2D mammograms and 2D slices derived from 3D DBT volumes. The model uses a Vision Transformer (ViT-base) architecture and was trained for 300,000 steps.
To evaluate its effectiveness, MammoDINO was tested on five major public mammogram benchmark datasets: RSNA, VinDr-Mammo, DDSM, CMMD, and CDD-CESM. It was assessed on various breast cancer screening tasks, including cancer detection, lesion detection, lesion type prediction, BI-RADS score prediction, and breast density classification. The performance was compared against several other models, including supervised CNNs (ResNet50, ConvNeXt), generic SSL encoders (DINOv2), radiology-specific SSL encoders (RadDINO), and weakly supervised medical vision encoders (BiomedCLIP, MammoCLIP).
The results showed that MammoDINO achieved state-of-the-art performance across most tasks and datasets. For example, on the VinDr dataset, MammoDINO scored 0.918 for cancer detection, outperforming all other models. Similar improvements were observed on DDSM, CDD-CSEM, and CMMD datasets. While it performed very well on RSNA, it was slightly behind in cancer detection and BI-RADS score prediction compared to some other models, but still highly competitive. An ablation study further confirmed that both the breast tissue aware augmentation and the 3D DBT adjacent slice loss significantly contribute to the model’s improved performance.
In conclusion, MammoDINO offers a scalable and annotation-free foundation for developing advanced computer-aided diagnosis tools for mammography. By learning clinically relevant features through its anatomically aware self-supervision, it can help reduce radiologists’ workload and improve the efficiency and accuracy of breast cancer screening. The principles behind MammoDINO are also adaptable and could potentially benefit other medical imaging modalities like CT and MRI.
Also Read:
- A New Method for More Reliable AI in Chest X-Ray Analysis
- Enhancing AI Reliability: A New Method for Detecting Unexpected Data
You can find more details about this research at the following link: MammoDINO Research Paper.
