Deep learning reconstruction of accelerated prostate MRI from image fidelity to diagnostic reliability
Deep learning reconstruction of accelerated prostate MRI from image fidelity to diagnostic reliability
Prostate MRI is time-consuming, which limits scanner capacity and increases waiting times. This thesis of Quintin van Lohuizen investigates whether prostate MRI can be accelerated using deep learning reconstruction without compromising the detection of clinically significant prostate cancer. Across four studies, accelerated scans were evaluated using automated detection models, region-specific image quality analysis, expert radiologist assessment, and uncertainty-based quality control. The results show that visually improved images do not necessarily preserve diagnostic information and that global image quality metrics can mask local degradation in diagnostically relevant regions. However, in a multi-reader study with eight expert prostate radiologists, detection of clinically significant prostate cancer was maintained up to sixfold acceleration, while ratings of sharpness, noise, and artifacts remained stable or improved. In addition, uncertainty maps generated during reconstruction provided useful quality control signals to identify scans that may require additional review.
Together, these findings indicate that substantially faster prostate MRI is feasible, but that safe implementation requires diagnostic endpoints, region-specific evaluation, and automated quality control. This work places accelerated MRI in a broader context of safe clinical AI deployment, showing that faster imaging is possible while maintaining diagnostic reliability and supporting responsible adoption of AI in clinical practice.