Synopsis

Keywords: Prostate, Machine Learning/Artificial Intelligence, Prostate cancer

In this prospective study, 56 male patients with suspected prostate cancer were included to evaluate an artificial intelligence (AI) based reconstruction method for T2-weighted sequences in multiparametric MRI (mpMRI). After comparison with conventionally acquired and reconstructed sequences we found the new AI-based method to produce images with higher image sharpness and delineation of lesions, confirmed by both qualitative and quantitative analysis. This is accompanied by a reduction of scan time by 29-37%. Confidence in the assessed PI-RADS scores was significantly higher for the AI-reconstruction. This new technique could therefore potentially increase diagnostic accuracy of mpMRI of the prostate.

Introduction

With an aging population fast and efficient MRI scans are needed to satisfy the increasing demand. Especially prostate cancer ranks among the most prevalent cancer types in men¹. An early and precise diagnosis by multiparametric MRI (mpMRI) is therefore not only crucial for long-term survival but also lays the foundation for the invasive procedure of a targeted biopsy². In recent years artificial intelligence (AI) has become one of the most promising methods in medicine; the field of radiology benefits greatly from these advances due to the high potential for image analysis³. The aim of this study was therefore to investigate qualitatively and quantitatively AI-enhanced T2-sequences in the setting of prostate mpMRI and whether it results in a better diagnostic confidence using the PI-RADS classification⁴.

Material and Methods

The prospective study was approved by the local review committee and all subjects gave written consent prior to mpMRI. Inclusion criteria were either an elevated prostate-specific-antigen of >4ng/ml or a suspicious digital rectal exam/transrectal ultrasound. All scans were performed on a 3 Tesla MRI (Philips Ingenia 3T). Besides T1- and diffusion-weighted sequences the scan protocol included a T2-weighted (T2w) cartesian low-resolution (T2_LR), a T2w cartesian high-resolution (T2_HR) and a T2w propeller (T2_PR) sequence. Derived from the raw data of the LR-sequence one additional sequence (T2_AI) was reconstructed using vendor provided prototype (Philips SmartSpeed Precise Image, SSPI). This AI-based reconstruction technique consists of a series of convolutional neural networks (CNNs): Adaptive-CS-Net⁵ allows to reconstruct images acquired with Compressed SENSE based variable density undersampling patterns. This CNN is applied prior to coil combination, removing the noise and undersampling artifacts from the images to obtain good image quality from accelerated acquisitions⁶. Subsequently, SSPI is an AI-model applied to replace the traditional zero-filling strategy and to increase the matrix size and therewith the sharpness of the images (super resolution)^7,8. This network is trained on pairs of low- and high-resolution data with k-space crops to induce ringing. The sequences were rated qualitatively by two radiologists in the categories artifacts, image sharpness, lesion conspicuity, capsule delineation and overall image quality as previously described⁹. Image sharpness was quantitatively assessed by measuring the apparent signal-to-noise (aSNR: SI_{peripheral zone}/SD_muscle) and contrast-to-noise ratio (aCNR: (SI_{peripheral zone}-SI_muscle)/SD_muscle), as well as the edge rise distance across the dorsal prostate capsule. The diagnostic confidence in PI-RADS scores for each sequence was rated separately on Likert-items from 1 (low) to 5 (high). Results are given as mean and standard deviation (continuous data) or median and interquartile range (ordinal data) as appropriate. Statistically analysis was conducted using Friedman test and one-way ANOVA. A P-value of <0.05 was considered as statistically significant.

Results

56 male patients were included in this ongoing study. Mean age was 67±8 years (range: 51–88 years) with a mean prostate-specific-antigen of 11.3±19.9 ng/ml. T2_AI was rated significantly better in comparison to the conventionally acquired sequences in the categories image sharpness, lesion conspicuity, capsule delineation and overall image quality (Tab. 1 and 2, Fig. 1 and 2). However, a difference in appearance of artifacts was not observed. Overall agreement between both raters was high with an intercorrelation coefficient of 0.862. Quantitative analysis (Fig. 3) showed a slightly higher aSNR for T2_AI (43.0±8.1) compared to T2_HR (42.0±8.3, P=0.009) and T2_PR (24.0±4.7, P<0.001), but not compared to T2_LR (43.1±8.1, P=0.54). The aCNR of T2_AI (38.2±7.9) was higher than all conventional sequences (T2_LR: 37.8±7.9, P=0.017; T2_HR: 35.9±8.3, P<0.001; T2_PR: 19.7±4.8, P<0.001). The edge rise distance as a quantifiable measurement of image sharpness was significantly lower with 0.824±0.379 mm for T2_AI (T2_LR: 1,117±0.412 mm, P<0.001; T2_HR: 1.445±0.789 mm, P<0.001; T2_PR: 1.176±0.797 mm, P=0.001). Acquisition duration of T2_LR, from which T2_AI is reconstructed, is reduced to 163±21s (T2_HR: 258±34s, P<0.001; T2_PR: 230±30s, P<0.001). Diagnostic confidence in PI-RADS scores was higher for T2_AI with a median of 4 [3-4] compared to T2_LR (3 [3-4], P=0.004), T2_HR (2.5 [2-3], P<0.001), and T2_PR (2 [2-3], P<0.001)(Fig. 2).

Discussion

Compared to the standard cartesian and propeller sequences, the new AI-reconstructed sequences show a significant improvement in nearly all aspects of image quality. This is both confirmed by a thorough qualitative and quantitative assessment. Furthermore, as the reconstruction of the AI-sequences starts from rapidly acquired low-resolution images, it effectively reduces the scan time between 37% (compared to T2_HR) and 29% (compared to T2_PR). However, the most important finding is the potential for an increased diagnostic performance, as a clear delineation of prostate lesions leads to a higher confidence in PI-RADS scores and therefore directly influences further diagnostics, most notably the decision for prostate biopsy. It must be noted in this context that we only investigated T2w-sequences and not diffusion-weighted sequences, which are equally important for the calculation of the PI-RADS score.

Conclusion

The evaluated AI-reconstructed T2w-sequences have high potential for significantly improving diagnostic workup of patients with prostate cancer by enhancing image quality, reducing scan time and resulting in a higher confidence of PI-RADS scores. In addition, this reconstruction technique is not limited to mpMRI and can be potentially applied to all MRI-sequences after initial image acquisition. However, a thorough investigation of different applications needs to be done to ensure high quality standards.

Acknowledgements

The present work was supported by Philips Healthcare.