Quantitative MRI or Machine Learning for Prostate MRI: Which Should You Use?

See also the article by Bonekamp et al in this issue.

Introduction

Prostate MRI is increasingly used to help locate clinically significant cancers for biopsy and therapy. Prostate MRI overcomes limitations of the “random” prostate biopsy performed via transrectal US that tends to overdetect indolent cancers while underdetecting clinically significant lesions. The use of prostate MRI has now been validated with large and randomized clinical trials, and some countries are incorporating it into their approved diagnostic pathways for prostate cancer diagnosis (1,2). However, concerns have been raised about the cost and variable quality of prostate MRI examination and whether the results obtained at academic centers will translate into community settings. One approach to the cost problem is to reduce the length of the examination by removing the dynamic contrast enhancement component. Instead, a 10–15-minute “biparametric” examination can be enabled by relying on the T2-weighted and diffusion-weighted components. While this reduces costs, it also makes interpretation more difficult. Nonetheless, reducing costs by shortening examination time seems like a good idea that is generally being applied to body MRI applications.

The quality issue has proven more difficult to solve as it relates not only to image quality but also to the quality of the interpretation. Solutions to the “quality” problem have naturally centered on training and education. However, it is difficult to expect a radiologist confronted with a growing number of specialized studies to perform at an expert level when interpreting a prostate MR image, especially if he or she rarely has encountered one. Both “artificial intelligence” and quantitative imaging have been proposed as potential solutions for helping less-experienced radiologists perform like an expert (3). In this issue of Radiology, Bonekamp et al compare qualitative radiologist interpretations of prostate MRI with a radiomic machine learning (RML) method and a quantitative method based simply on apparent diffusion coefficient (ADC) measurements (4). They show superiority of both methods over the conventional interpretation, but no difference between RML and ADC. At first read, this is very welcome news as it suggests two avenues, one of which is quite simple, to improve the quality of prostate MRI.

Quantitative MRI has long been a holy grail of imagers who dream that one day MRI could be performed in such a uniform way that reliable measurements of MR parameters could be obtained across different centers. To date, most quantitative methods in MRI have failed because they require too much time (eg, measuring T2-weighted values), there is too much physiologic and methodologic variability (eg, measuring dynamic contrast enhancement parameters such as volume transfer constant, or K^trans), or they are simply not that valuable. Generations of MR physicists have fought the good fight to quantitate MRI; however, manufacturers continue to make proprietary software and hardware tweaks that make standardization nearly impossible. This renders comparisons of one MRI vendor to another (and even one software upgrade to another within the same vendor) very difficult. The only measurement that has withstood these assaults is ADC, which is fairly robust across platforms. Therefore, it is very satisfying to see Bonekamp et al report that ADC measurements outperformed qualitative assessments and were at least equal to machine learning.

However, while physicists have been waging a heroic battle to introduce quantitation into MRI, machine learning is fast emerging as a method of identifying abnormalities on images. Machine learning employs a bewildering number of parameters, most of which have no direct real-world meaning. These parameters are mysteriously combined into “random forest clusters” and probability maps emerge. For someone steeped in the tradition of explaining why each MR sequence is important based on physiology (eg, edema, water diffusion, vascularity), the mash-up of parameters in machine learning is intellectually unsatisfying. For instance, in this study the machine learning algorithm combined a total of 846 image features across three MRI sequences, encompassing nebulous concepts such as first-order features, volume, shape, and texture. Thus, it is undeniably satisfying when Bonekamp et al conclude that a single ADC measurement performs as well as the somewhat inscrutable radiomic machine learning.

However, it may be a bit early to celebrate a victory for quantitative MRI. While ADC values are likely the most robust of the MR parameters, they do vary from site to site and machine to machine. This data set was obtained on a single 3-T state-of-the-art MRI unit from one vendor at one center of excellence. This is a far cry from reality, in which vendors, field strengths, and most of all, expertise, vary tremendously across hospitals. Meanwhile, we are still in the early days of machine learning. Experts tell us that an effective machine learning algorithm requires between 1000 and 10 000 pieces of data for robust training. Only 183 patients were used here to train the algorithm. Imagine what it could do with more training. Moreover, the training occurred against MRI–transrectal US fusion biopsy results and not radical prostatectomies. Therefore, the false-negative rate is hard to determine as one would not biopsy what one cannot see (5). It is likely that machine learning will excel at extracting features that are either not visible or barely visible to the human observer (6). It will be challenging to obtain carefully curated data sets of prostate MR images and the corresponding annotated pathology; however, eventually this will happen. When it does, it is likely that we will be amazed by how good machine learning is at augmenting the human reader.

While the results of Bonekamp et al are a satisfying rebuttal to the overwhelming hype surrounding machine learning in imaging, it is probably only a temporary respite. Despite substantial impediments, there are also tremendous forces in its favor, including powerful and wealthy IT companies such as IBM and Alphabet, along with the great equalizing force of data clouds, high speed networks, and most of all, brilliant young minds born into the digital age. As someone who has labored long and hard to standardize MRI for quantitation, I have to confess my bet is on machine learning.