Enhancing deep learning methods for brain metastasis detection through cross-technique annotations on SPACE MRI

Area of Science:

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

Background:

• Gadolinium-enhanced Sampling Perfection with Application-Optimized Contrasts Using Different Flip Angle Evolution (SPACE) sequences offer superior visualization of brain metastases (BMs) compared to Magnetization-Prepared Rapid Acquisition Gradient Echo (MPRAGE) sequences.
• The enhanced conspicuity of BMs on SPACE images can potentially lead to higher-quality annotations (HAQ), which may improve the performance of deep learning (DL) algorithms for BM detection on MPRAGE images.

Purpose of the Study:

• To investigate whether high-quality annotations (HAQ), derived from SPACE MRI sequences, can enhance the performance of deep learning (DL) algorithms for detecting brain metastases (BMs) on standard MPRAGE images.
• To evaluate the impact of cross-technique transfer learning, using HAQ from SPACE images, on the detection and delineation accuracy of DL models.

Main Methods:

• A retrospective analysis of contrast-enhanced MRI data from 157 patients with BMs was performed, including both SPACE and MPRAGE sequences.
• Two annotation strategies were employed: normal annotation quality (NAQ) on MPRAGE and high-quality annotation (HAQ) on coregistered SPACE images.
• Multiple DL models were trained using either NAQ or HAQ, with images from either MPRAGE or SPACE sequences, and evaluated on internal and external test datasets (660 patients) for detection and delineation performance.

Main Results:

• DL models trained with HAQ demonstrated superior performance compared to those trained with NAQ.
• The SPACE-HAQ model achieved a positive predictive value (PPV) of 0.978, sensitivity of 0.882, and F1-score of 0.916.
• Relative to MPRAGE-NAQ, the MPRAGE-HAQ model showed significant improvements in F1-score (2.5–9.6 points, p < 0.016) and sensitivity (4.6–8.5 points, p < 0.001) on additional test datasets, with volumetric instance sensitivity also improving (3.6–7.6 points, p < 0.001).

Conclusions:

• High-quality annotations (HAQ), even when derived from a different sequence (SPACE) than the target application (MPRAGE), significantly improve the performance of deep learning models for brain metastasis detection.
• This cross-technique transfer learning approach can enhance DL model accuracy without requiring specialized imaging sequences at the time of application, offering a practical method for improving diagnostic performance.
• HAQ alone accounts for approximately 40% of the performance gains achieved with SPACE images as input, enabling accurate and automated detection of small BMs (<1 cm).