Breast Tumor Identification in Ultrafast MRI Using Temporal and Spatial Information.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Distinguishing benign from malignant breast lesions is crucial for effective patient management.
• Standard dynamic contrast-enhanced MRI (DCE-MRI) is a common imaging technique, but its accuracy can be limited.
• Ultrafast MRI offers potential for improved lesion characterization by capturing rapid temporal dynamics.

Purpose of the Study:

• To evaluate the feasibility of deep learning models for differentiating benign and malignant breast lesions using ultrafast MRI.
• To assess the combined value of spatial and temporal information from ultrafast MRI in breast lesion characterization.

Main Methods:

• Retrospective analysis of 173 breast lesions from 122 women using ultrafast MRI.
• Development of a 2D convolutional neural network (CNN) for spatial feature extraction and a long short-term memory (LSTM) network for temporal feature extraction.
• Performance evaluation using 100-times repeated stratified four-fold cross-validation and comparison with standard DCE-MRI models.

Main Results:

• Ultrafast MRI-based 2D CNN achieved a mean AUC of 0.81 ± 0.06; LSTM achieved 0.78 ± 0.07.
• The combined CNN-LSTM model demonstrated a superior mean AUC of 0.83 ± 0.06.
• Ultrafast MRI models significantly outperformed standard DCE-MRI models in malignancy discrimination.

Conclusions:

• Deep learning models utilizing ultrafast MRI show enhanced performance in discriminating malignant breast lesions.
• The integration of temporal information via LSTM networks provides added value for breast lesion characterization.
• Ultrafast MRI combined with deep learning represents a promising advancement in breast cancer diagnostics.