Performance of an ultra-fast deep-learning accelerated MRI screening protocol for prostate cancer compared to a standard multiparametric protocol

We evaluated the diagnostic performance and quality of an ultra-fast DL accelerated screening protocol in biopsy-naïve men compared to a standard mp protocol as used in the current clinical routine. The abbreviated bp protocol was considered diagnostic in all patients and showed no significant difference in diagnostic performance compared to the mp protocol. Furthermore, the image quality of the abbreviated protocol was rated as better by both readers. Those results are in line with several studies which evaluated DL-sequences for MRI of the prostate, although neither evaluated an entirely DL-accelerated screening protocol but merely individual DL-accelerated sequences or DL-accelerated sequences as part of a conventional protocol [26‐29].

Considering the standard mp protocol as the reference standard, both sensitivity and specificity were excellent for R1. These results are also reflected by Kuhl et al who found similar diagnostic accuracy for abbreviated bpMRI and mpMRI. Only one additional csPCa was detected in mpMRI compared to bpMRI as opposed to 11 additional false-positive diagnoses [30]. Van der Leest et al found that an abbreviated biparametric protocol resulted in a surplus of 2% in biopsies and 1% overdetection of low-grade PCa compared to the standard protocol with similar sensitivity (95%) and a slightly lower specificity (65% compared to 69%) in 626 patients [31].

Employing a cut-off of PI-RADS ≥ 3, R2 achieved an equally high sensitivity at the cost of specificity (98%/83%). An adverse effect was observed regarding a threshold of PI-RADS ≥ 4, lacking sensitivity instead (89%/98%). This effect may be due to slightly less experience in reporting prostate MRI [32]. Coherently, this is reflected in inter-reader agreement, showing only substantial agreement between readers at most (κ: 0.61), whereas intra-reader agreement was substantial to excellent. This implies that different PI-RADS scores between readers most likely emerge from varying interpretations of lesion characteristics rather than from differences in image quality. Additionally, we observed an unequal distribution of PI-RADS scores in both the abbreviated and the mp protocol for both readers (refer to Table 3). In particular, barely any PI-RADS 1 lesions and only a few PI-RADS 5 lesions were included. This limits generalizability for PI-RADS 1 lesions but implies robust diagnostic performance for PI-RADS 2–4 lesions. These lesions tend to be most debatable in clinical practice due to the possibility of an upgrade in case of positive DCE or higher DWI categories. Heterogeneity in the distribution of PI-RADS scores between readers was higher in the mp protocol. This finding was to be expected as the mp protocol offers more options to inspect the lesion in its three-dimensional shape and to upgrade scores via DCE.

We consider that enhanced image quality is mainly due to the reconstruction of the DL-algorithm, but also critically related to the reduced acquisition time, resulting in less motion artefacts [33]. We used the resolve DWI as a reference, as it serves as our routine sequence for prostate exams. To find a shorter alternative sequence a single shot EPI technique was chosen, which is more time efficient compared to the segmented readout resolve acquisition. By integrating a DL reconstruction in the EPI sequence scan time could be even further cut down. Besides the reduction of motion artefacts another potential benefit of the reduction of scan time is an increase in patient comfort and compliance, especially valuable in an elder patient collective as laying still becomes an increasing problem during the examination period [34, 35]. Abbreviated protocols are especially suitable for specific indications that do not require extensive three-dimensional information or contrast enhancement, e.g., screening for PCa. This offers higher MRI throughput with increased expediency at a potentially lower cost per examination, given the indication is chosen correctly [36]. However, single-planar protocols may not suffice for a more detailed anatomical evaluation of the tumor, potential extracapsular growth, or local staging [37, 38]. Adequate resolution in a single axial plane already resolved issues of volume measuring or co-registration of the gland for targeted biopsy in our institution, although abbreviated protocols have yet to be evaluated thoroughly for tumor segmentation prior to local therapy, follow-up examinations and assessment of local tumor recurrence [39]. Another disadvantage of providing more than one standard protocol or even numerous variants is a potential increase in the complexity of protocol selection within the clinical workflow, possibly resulting in inadequate examinations [40]. Furthermore, indefinite results of abbreviated protocols and recall of patients for additional examinations may limit their acceptance. Consequently, in order to exploit the advantages of shorter scan times, clinical workflows, as well as indications for examinations need to be further refined [41].

Although carried out prospectively, the study was conducted in a single centre setting with a relatively small cohort. Furthermore, seven patients were excluded due to incomplete scans and four patients due to severe artefacts. While effects from incomplete scans represent a random error and artefacts affected both the abbreviated and the multiparametric protocol and effects, this might slightly distort the initial cohort and limit generalizability further. Despite the fact that these initial results suggest a benefit in image quality and imaging efficiency at an adequate diagnostic performance, external validation is required, ideally in a multi-center setting.

As we compared slightly different sequence types of DWI (readout-segmented DWI, in this case, RESOLVE as currently implemented standard vs. ZOOMit single shot EPI), direct comparability might be limited. In the ZOOMit sequence, regions outside of the field of view do not contribute to the gathered MR signal, therefore preventing infolding of signals outside of the field of view. Consequently, motion artefacts are reduced to a minimum [42]. A recent study by Klingebiel et al suggests sufficient comparability as they showed that zoomed sequences outperformed conventional multi-shot readout sequences regarding objective image quality [43].

Omitting DCE sequences in the abbreviated protocol might hinder the upgrade of PI-RADS 3 scores to PI-RADS 4 in the peripheral zone compared to the multiparametric protocol. However, recent evidence suggests limited benefit of upgrading those lesions as they seem to yield significantly less csPCa compared to PI-RADS 4 lesions that were not upgraded (14.5% csPCa in PI-RADS 3 with positive DCE and 53.3% csPCa in PI-RADS 4 lesions) and decrease detection rates for csPCa [4]. The authors suggest reconsidering the current scoring categorization and propose to set the indication for biopsy according to the PSAD for PI-RADS 3 lesions with positive and negative DCE alike. These findings generally question the implementation of DCE imaging in PI-RADS scoring but even more so justify the omission in a screening setting as a tradeoff for efficiency.

Furthermore, we compared the abbreviated protocol to the current gold standard for imaging of PCa (mpMRI) without histopathological confirmation, so we could only evaluate diagnostic performance but not diagnostic accuracy. Although we found no significant difference between the abbreviated and the standard protocol and similar results can be expected, validation studies are desirable. Currently, there is a general lack of evidence in the comparison of abbreviated and standard protocols, requiring further research. Moreover, additional investigation of implementation and practicability in clinical routine and cost-effectiveness will be necessary.

In conclusion, an abbreviated screening protocol for prostate cancer in biopsy-naïve men proved similar diagnostic performance and better imaging quality compared to the conventional multiparametric protocol, requiring less than 15% of scan time. This results in a major improvement in imaging efficacy, enabling the implementation of screening protocols in clinical routine.