Improving the diagnostic quality of fetal MRI: Evaluation of an AI-based post processing approach for ultrafast image enhancement

Aim and Research Question(s)

This thesis evaluates an AI-based super-resolution algorithm to improve image quality in accelerated T2-weighted fetal MRI. The study investigates if this post-processing technique can (1) enhance ultrafast images, (2) offer diagnostic value comparable to high-resolution scans, (3) facilitate shorter acquisition times, and (4) meet the standards of expert radiologists.

Background

Fetal MRI is challenged by unpredictable fetal motion, which necessitates ultrafast acquisition protocols. However, this speed comes at the cost of lower image quality. AI-based super-resolution (SR) offers a potential solution by retrospectively enhancing accelerated images [1].

Methods

Within an ethically approved study (EK Nr: 2028/2023), 30 retrospective fetal MRI cases were analyzed. Three T2-weighted TSE sequences were compared: T2 TSE cs: Conventional sequence, accelerated with compressed sense (CS), T2 AI SR: The same accelerated sequence with AI post-processing, T2 HR cs: Standard time-intensive High Resolution sequence. Quantitative Analysis (N=30/20): Image quality was objectively assessed using SNR, CNR, and ERD. Statistical comparison were made using the Wilcoxon signed-rank test. Qualitative Analysis (N=11): Two expert neuroradiologists performed a blinded, randomized evaluation of image quality using a 4-point Likert scale (1=Excellent, 4=Non-diagnostic) for 5 criteria, including Artifacts, Tissue Contrast, Image Sharpness, Overall Image Quality and Diagnostic Confidence.

Results and Discussion

Qualitative Expert Evaluation: The AI sequence was rated significantly superior to both Conventional and HR sequences in Image Sharpness, Tissue Contrast, Overall Image Quality, and Diagnostic Confidence (all p<0.001). Expert noted marked improvement in the delineation of fine anatomical structures. Quantitative Metrics: No significant improvement in SNR or CNR was observed for the AI sequence. Traditional metrics did not align with the qualitative findings, suggesting they may not fully capture the perceptual benefits of AI-driven reconstruction. Scan Time: Superior perceived quality was achieved in 38% less time with the AI sequence compared to the HR reference (34.5s vs. 56s).

Conclusion

AI-based post-processing significantly improves the perceived diagnostic quality of ultrafast fetal MRI, outperforming even slower HR protocols in expert evaluations. This allows for a substantial reduction in scan time without compromising diagnostic confidence. The study highlights that human-centered expert validation is indispensable for assessing advanced AI tools, as traditional quantitative metrics may not reflect their true clinical utility. 

References

[1]L. M. Bischoff u. a., „Deep Learning Super-Resolution Reconstruction for Fast and Motion-Robust T2-weighted Prostate MRI“, Sep.2023, doi: 10.1148/radiol.230427.