TLDR: A new study evaluates an AI model, the Medical Slice Transformer (MST), for quickly identifying suspicious breast lesions (BI-RADS ≥4) in MRI scans, including both contrast-enhanced and non-contrast-enhanced protocols. The model achieved 19% specificity for contrast-enhanced and 17% for non-contrast-enhanced MRI at 97.5% sensitivity, showing promise for improving breast cancer screening efficiency by helping to triage cases that are likely benign.
Breast cancer detection relies heavily on Magnetic Resonance Imaging (MRI) due to its high sensitivity, especially for women with dense breast tissue where mammography might be less effective. However, interpreting breast MRI scans is a time-consuming process that requires significant expertise from radiologists. This challenge often limits the widespread implementation of breast MRI screening, despite its proven value in early detection.
To address this, researchers have been exploring the potential of Artificial Intelligence (AI) to assist in pre-screening and triaging breast MRI examinations. A recent study introduces an adapted foundation model, the DINOv2-based Medical Slice Transformer (MST) framework, designed to help identify cases that are highly likely to be benign, thereby streamlining the diagnostic workflow.
The primary goal of this research was to evaluate the MST framework’s ability to effectively rule out significant findings (classified as Breast Imaging Reporting and Data System, or BI-RADS, ≥4) in both contrast-enhanced and non-contrast-enhanced abbreviated breast MRI protocols. Abbreviated protocols are shorter MRI sequences that aim to reduce examination time, cost, and the need for contrast agents, which carry rare potential health risks.
The study involved a retrospective analysis of 1,847 single-breast MRI examinations from an in-house dataset and an additional 924 from an external validation dataset (Duke). Four different abbreviated protocols were tested: T1-weighted early subtraction (T1sub), diffusion-weighted imaging with a b-value of 1500 s/mm² (DWI1500), a combination of DWI1500 and T2-weighted imaging (DWI1500+T2w), and a combination of T1sub and T2-weighted imaging (T1sub+T2w).
Key Findings
The MST model demonstrated promising results. The combination of T1sub+T2w achieved the highest Area Under the Receiver Operating Characteristic Curve (AUC) of 0.77. For non-contrast-enhanced imaging, DWI1500+T2w showed an AUC of 0.74, with no statistically significant difference compared to the contrast-enhanced protocols. This suggests that non-contrast approaches could be viable for screening.
Crucially, at a high sensitivity threshold of 97.5% (meaning the model correctly identified 97.5% of suspicious lesions), the T1sub+T2w protocol achieved a specificity of 19% for contrast-enhanced MRI, while DWI1500+T2w achieved 17% specificity for non-contrast-enhanced MRI. Specificity indicates the model’s ability to correctly identify truly benign cases. This means the model could effectively triage a significant portion of benign cases, reducing the workload for radiologists.
The analysis of false negative cases (suspicious lesions missed by the model) revealed that these were predominantly small, often non-mass enhancements, with a mean diameter typically less than 10 mm at higher sensitivity thresholds. This highlights areas where the model could be further improved.
External validation on the Duke dataset, using the T1sub-trained model, also yielded an AUC of 0.77, indicating good generalization capabilities. Furthermore, the explainability of the model was assessed through attention maps, which showed that in 88% of true positive cases, the model correctly focused on the relevant imaging features, providing confidence in its decision-making process.
Also Read:
- New AI Tool Detects Unrealistic Shapes in Synthetic Medical Images
- Advancing Cancer Diagnosis: A New End-to-End Approach for Whole Slide Image Analysis
Future Outlook
While the MST framework shows significant potential for triaging breast MRI examinations, the study emphasizes that further research is needed before clinical implementation. The performance levels, while promising for an AI tool, are still lower than those of human radiologists in comparable screening scenarios. However, the ability of the model to effectively rule out cases without significant findings at high sensitivity suggests a viable pathway for optimizing breast MRI screening approaches, potentially improving early detection and patient outcomes by making the process more efficient and accessible. The source code for this model is publicly available, encouraging further development and research. You can read the full research paper for more details on the methodology and results. Read the full research paper here.
