AI-driven diffusion weighted imaging-based non-contrast protocol for breast cancer diagnosis: a multicentre,...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Contrast-enhanced breast MRI is sensitive but complex, time-consuming, and uses contrast agents.
• Noncontrast diffusion-weighted imaging (DWI) is fast but lacks diagnostic accuracy for standalone use.
• Investigating a deep learning (DL) model to improve DWI accuracy for breast cancer diagnosis.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for breast cancer diagnosis using only diffusion-weighted imaging (DWI).
• To compare the diagnostic performance of the DWI-DL model against abbreviated enhanced-based (AE-DL) models and expert radiologists.
• To evaluate the clinical utility of the DWI-DL model in a selective contrast-enhanced workflow.

Main Methods:

• A DWI-based DL model (DWI-DL) was developed using data from 1286 patients.
• Independent testing involved three external cohorts (n=661) and a prospective cohort (n=546).
• Multireader multicase validation assessed the AI-guided selective sequence protocol's performance and interpretation time.

Main Results:

• The DWI-DL model showed diagnostic performance comparable to the AE-DL model across all cohorts (AUC: 0.771-0.912 vs. 0.780-0.898).
• DWI-DL outperformed expert radiologists interpreting DWI alone (AUC: 0.781-0.858 vs. 0.714-0.770).
• The AI-guided selective protocol was non-inferior to the full protocol (AUC: 0.834 vs. 0.835) and reduced interpretation time by 55.5%.

Conclusions:

• A DL model utilizing only noncontrast DWI can accurately diagnose breast cancer.
• The DWI-DL model offers robust performance, surpassing human DWI interpretation and reducing time.
• The DWI-DL model combined with selective contrast-enhanced sequences presents a promising tool to streamline breast cancer diagnosis.