Assessing deep learning reconstruction for faster prostate MRI: visual vs. diagnostic performance metrics.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnostics

Background:

• Deep learning (DL) MRI reconstruction allows faster scanning with good visual quality.
• Assessing the diagnostic impact of DL reconstruction often requires extensive reader studies.
• Existing diagnostic DL tools can be leveraged to evaluate reconstructed image quality.

Purpose of the Study:

• To assess the diagnostic quality of DL-reconstructed MRI images using existing diagnostic DL.
• To compare the visual and diagnostic performance of DL reconstruction against standard methods.
• To determine the clinical relevance of DL reconstruction in prostate cancer detection.

Main Methods:

• Retrospective analysis of 1535 biparametric prostate MRIs (2016-2020).
• Expert radiologists delineated clinically significant prostate cancer (csPCa) lesions.
• T2-weighted scans were undersampled, and DL reconstruction (DLRecon) and DL detection (DLDetect) were applied.
• Evaluated partial area under the Free-Response Operating Characteristic (pAUC FROC) curve and structural similarity (SSIM).

Main Results:

• DLRecon improved visual quality (SSIM) at 4- and 8-fold undersampling.
• Diagnostic performance (pAUC FROC) for DLRecon was significantly lower than fully sampled scans.
• Similar trends of reduced diagnostic performance were observed at higher undersampling rates.

Conclusions:

• DL reconstruction yields visually appealing MRI images but may compromise diagnostic accuracy.
• The use of DL reconstruction in clinical settings requires caution due to reduced cancer detection rates.
• Integrating diagnostic AI provides essential metrics for adopting DL reconstruction models into practice.