A Deep Learning-Based Approach to Reduce Rescan and Recall Rates in Clinical MRI Examinations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic resonance (MR) imaging rescans and recalls incur substantial hospital revenue loss.
• Motion artifacts in brain MR series are a primary cause for rescans and recalls.

Purpose of the Study:

• To develop a rapid, automated method for assessing the need for rescans in motion-corrupted brain MR series.
• To reduce financial losses associated with unnecessary MR rescans and recalls.

Main Methods:

• A deep learning (DL) model was developed to output a probability score indicating the clinical usefulness of an MR series.
• The DL model's performance was compared against assessments from technologists and radiologists using 49 test series with varying motion artifacts.
• The DL model's classification was evaluated against thresholds adjusted for scan indication and reading radiologist.

Main Results:

• Image quality ratings were dependent on scan indication and the reading radiologist.
• The DL algorithm demonstrated a statistically significant decrease in recall rates for multiple sclerosis (MS) screening (P = .03).
• The DL approach achieved rescan/recall ratios of (7.3 ± 2.2)/(3.2 ± 2.5) for MS and (3.6 ± 1.5)/(2.8 ± 1.6) for stroke indications.

Conclusions:

• Automated deep learning-based image quality assessment can significantly decrease rescan and recall rates.
• This AI-driven method provides technologist-independent image quality ratings.
• Implementing this DL approach could save hospitals an estimated $24,000 per scanner annually by reducing rescans and recalls.