Impact of deep learning image reconstruction on ADC quantification and histogram metrics: a phantom study

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• Deep learning (DL) reconstruction enhances magnetic resonance imaging (MRI) quality and reduces scan time.
• The impact of DL on apparent diffusion coefficient (ADC) quantification in diffusion-weighted imaging (DWI) is not fully understood.

Purpose of the Study:

• To investigate the influence of DL-based reconstruction on ADC quantification and histogram metrics.
• To assess the accuracy and repeatability of ADC values derived from DL-reconstructed DWI data.

Main Methods:

• A calibrated DWI phantom with known ADC values was scanned using conventional (DL-OFF) and varying DL strength levels.
• Both full (fFOV) and reduced (rFOV) field-of-view sequences were employed.
• ADC quantification, repeatability (CV), accuracy, and histogram features were analyzed.

Main Results:

• DL reconstruction demonstrated high ADC accuracy (-2% to 7%) and repeatability (0.1% to 1.2% CV).
• DL progressively reduced ADC histogram dispersion, particularly in high-ADC vials, without altering median ADC values.
• Significant decreases in entropy and interquartile range were observed with increasing DL strength.

Conclusions:

• DL-based reconstruction preserves ADC accuracy and repeatability in DWI, even with reduced field-of-view.
• DL narrows ADC distributions, reducing dispersion and showing potential for oncologic imaging.
• Further validation in clinical settings is needed to assess the generalizability of DL reconstruction in DWI.