Deep learning-accelerated T1-weighted pelvic imaging may shorten prostate MRI while preserving image quality and perceived diagnostic confidence for metastasis screening. A prospective feasibility evaluation in European Radiology Experimental compared research deep learning-accelerated T1-weighted VIBE Dixon sequences with a conventional T1-weighted VIBE Dixon sequence in patients undergoing prostate MRI. The work focused on image quality and reader confidence, not formal sensitivity or specificity for lymph node or bone metastasis detection. The accelerated sequences reduced acquisition time from 143 seconds to either 45 seconds or 15 seconds, while maintaining excellent perceived diagnostic confidence.
Deep Learning Shortens Pelvic T1 Acquisition
Prostate MRI includes T1-weighted imaging for all examinations under PI-RADS version 2.1. T1-weighted imaging can help detect intraprostatic haemorrhage and screen for pelvic metastases, particularly after contrast administration. The Dixon technique supports fat-water separation and can aid characterisation of bone lesions through improved fat suppression and contrast resolution.
The evaluation included consecutive patients referred for multiparametric prostate MRI between February and April 2024 for clinical suspicion of prostate cancer or assessment of known prostate cancer without prior surgical treatment. After exclusions for incomplete acquisition of the research sequences, 54 patients remained in the final analysis. MRI indications included cancer detection, primary staging and active surveillance.
All patients underwent prostate MRI at 3 T with T2-weighted imaging, diffusion-weighted imaging and dynamic contrast-enhanced imaging, in accordance with PI-RADS version 2.1. In addition to the conventional T1-weighted VIBE Dixon sequence, two research deep learning-accelerated versions were acquired before and after gadolinium-based contrast administration. The standard sequence used an acquisition time of 143 seconds. The accelerated versions reduced acquisition time to 45 seconds and 15 seconds.
The deep learning-enhanced reconstruction used undersampled k-space data and coil sensitivity maps, followed by neural network-based image enhancement and deep learning-based super-resolution. Water and fat images were then generated with the scanner-integrated Dixon algorithm.
Image Quality Remains High Across Sequences
Three radiologists with experience in prostate MRI independently assessed overall image quality, sharpness, subjective noise, artefacts and perceived diagnostic confidence. The readers were blinded to sequence type and clinical data. Images were presented in random order, and no prior familiarisation session with deep learning images was conducted.
Must Read: AI Model Improves Prostate MRI Distortion Correction
For non-contrast imaging, overall image quality was good across all three sequences. The 45-second deep learning sequence showed improved sharpness and reduced noise compared with the standard sequence. The 15-second sequence also showed favourable image quality, although its ratings were slightly lower in some non-contrast assessments. More artefacts were noted in both deep learning-accelerated non-contrast sequences than in the standard sequence, with ratings distributed mainly between mild and minimal artefacts.
Perceived diagnostic confidence for pelvic lymph nodes and bone lesions remained excellent across all non-contrast sequences. Assessment of intraprostatic haemorrhage showed high scores overall, but the 15-second non-contrast sequence reached a slightly lower median rating than the standard and 45-second accelerated sequences.
For contrast-enhanced imaging, both accelerated sequences outperformed the standard sequence in overall image quality. The 15-second contrast-enhanced sequence reached an excellent median quality score. Sharpness improved most clearly with deep learning acceleration, while subjective noise was lower than with the standard contrast-enhanced sequence. The 45-second contrast-enhanced sequence showed more artefacts than the standard and 15-second contrast-enhanced sequences. Despite these differences, perceived diagnostic confidence for lymph nodes and bone lesions remained excellent across all contrast-enhanced sequences.
Quantitative Findings and Limits Shape Interpretation
Quantitative assessment showed low variation of signal intensity in the piriformis muscle across all sequences, indicating overall homogeneous signal. Both deep learning-accelerated sequences had minimally more inhomogeneous signal than the standard sequence, in both non-contrast and contrast-enhanced imaging. The absolute differences were small.
Additional texture analysis showed lower mean signal intensity in the accelerated sequences, while within-region signal standard deviation was comparable across sequences. Skewness and kurtosis showed no clinically relevant differences. Entropy was slightly higher in the deep learning-accelerated non-contrast sequences, suggesting mildly increased texture complexity. These findings indicate that the accelerated reconstruction altered some image characteristics while largely preserving signal distribution properties.
Non-inferiority analysis confirmed that both accelerated sequences were non-inferior to the standard sequence for diagnostic confidence regarding lymph nodes and bone lesions in native and contrast-enhanced imaging. For intraprostatic bleeding assessment in native sequences, the 15-second accelerated sequence met non-inferiority criteria. The 45-second sequence did not meet these criteria, with a subset of patients showing deterioration of at least one Likert point.
The evaluation has several limitations. The patient group was modest, and the number of patients with metastases was too limited for statistical testing of diagnostic performance. The work assessed image quality rather than sensitivity, specificity, clinical outcomes or patient management. The algorithm was vendor-specific, and independent external validation for pelvic imaging had not been performed. The results apply only to T1-weighted VIBE Dixon imaging and may not transfer to other sequences or clinical indications.
Research deep learning-accelerated T1-weighted VIBE Dixon imaging appears feasible for prostate MRI, with markedly shorter acquisition time and preserved perceived diagnostic confidence for pelvic lymph node and bone lesion assessment. Image sharpness improved, while artefacts and signal inhomogeneity increased slightly in some accelerated sequences. The findings support further evaluation of ultrafast T1-weighted pelvic imaging as a time-saving component of prostate MRI, while larger studies remain necessary to assess diagnostic performance, clinical impact and transferability beyond the evaluated vendor-specific implementation.
Source: European Radiology Experimental
Image Credit: iStock
References:
Nedelcu A, Russe MF, Wilpert C et al. (2026) Feasibility of deep learning-accelerated ultrafast T1-weighted VIBE Dixon imaging of the pelvis for screening of metastases in prostate MRI. Eur Radiol Exp; 10, 98.
