Can MRI Plus Artificial Intelligence Diagnose Prostate Cancer?
The question posed in the title is a worthy one, given the increasing attraction of applying Artificial Intelligence (AI) to medicine, specifically in detecting and diagnosing prostate cancer (PCa). Why do we need this? Is there something lacking in the way doctors have been performing? Well, not exactly.
Radiologists and urologists who are experienced in diagnosing PCa do a great job, but they can potentially do even better when a) they are equipped with the most advanced methods and technologies, b) they have AI assistance, and c) they are not burdened with heavy workloads.
New study addresses both improvements
We have breaking news! As of this writing, a pre-publication study by Youn, et al. (Aug. 2021) has been released online, pending final peer review edits.[i] It’s an exciting study that contributes insights into both advanced MRI/AI technology, as well as a way to expedite diagnosis and reporting for busy radiologists. Patient and urologist demand for MRI of the prostate has increased. A heavy workload of capturing, reading, interpreting and reporting MRI results can slow down even the most experienced radiologist.
The Youn study is a comparison between an AI deep learning-based algorithm (DLA) vs. radiologists with various experience levels from least to most as follows:
- Reader level 1 – Group of four 2-year residents
- Reader level 2 – Group – A of four 3-year residents
- Reader level 3 – A board-certified radiologist with 4.5 years’ experience
- Reader level 4 – A board-certified radiologist with 5 years’ experience
- Reader level 5 – A board-certified radiologist with 10 years’ experience.
The DLA used for the comparison had been trained using 2,170 biparametric MRIs (bpMRI means T2 weighted + diffusion weighted imaging sequences) to identify the location of up to 5 suspicious lesions visible on each individual scan, and classify the probability of PCa using the 5-point PI-RADS scoring system. NOTE: For study purposes, the authors used Version 2, called PI-RADS v2; a more recent version, PI-RADS v2.1, is worded with greater imaging, interpretation and reporting detail.
In terms of advanced technology, all research images were obtained on a state-of-the-art 3T (3 Tesla) magnet. Furthermore, the DLA developed by the magnet manufacturer is in testing and not yet commercially available, but is on the cusp of AI’s future in image-based detection and diagnosis.
Study design and methods
The study authors used 121 PCa patient cases; 58 patients had undergone pre-biopsy biparametric MRI (bpMRI has T2 -weighted plus diffusion-weighted imaging sequences), while 63 had undergone multiparametric MRI (mpMRI has bpMRI plus at least one additional sequence). Other case information included patient age, PSA, medication use (5-alpha reductase inhibitors such as Proscar, Avodart, etc.), pathology reports from biopsy (and prostatectomy if applicable), clinical reports and other imaging/biopsy details.
For the study, only bpMRI scans were used; bpMRI data can be extracted from mpMRI scans, so all images were apples-to-apples. All images were analyzed radiologists and DLA. Each reader level recorded the number, location and size of suspicious PCa lesions; DLA likewise identified up to 5 suspicious lesions per MRI. All readers and DLA classified each lesion according to PI-RADS, probability of cancer ranging from 1 (very low) to 5 (very high), and identified the index lesion by PI-RADS and size.
The image-based findings were correlated with pathology reports (physical PCa diagnosis) and clinical reports. Then, statistical analysis was conducted to compare “the diagnostic performance of per-patient PI-RADS scores for all readings…”, that is, clinical reports, the five radiologists [levels], and DLA, and to “calculate the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy in detecting prostate cancer…” for each.
Findings and conclusion
Pathology (biopsy/prostatectomy) confirmed clinically significant PCa in 43 patients (35.5%). Not surprisingly, among the radiologists, diagnostic accuracy and PI-RADS classification varied according to experience. The more experience, the greater the accuracy. Among key findings:
- “The sensitivity and specificity of the DLA and the expert were similar in detecting clinically significant prostate cancer for a PI-RADS cutoff value ≥ 4.”
- “The performance of DLA was similar to that of clinical reports from various radiologists in clinical practice.”
While the DLA was significantly more accurate than reader level 1, it was significantly lower than reader level 5. However, as explained above, it was similar to the other reader levels as well as clinical reports.
The authors concluded that “…the moderate agreement between DLA and the expert radiologist seems promising and DLA-based PI-RADS categorization may help to improve inter-reader variability in clinical practice.” For practical purposes, AI-based lesion detection and classification may help reduce radiologists’ review and interpretation time by generating reports which the human reader can edit and have final determination. This can be of great assistance for less experienced readers. Since this study is “the first comparison between DLA and radiologists with various levels of experience in PI-RADS classification,” it is a tremendous contribution to affirming the merits of applied AI in PCa diagnosis.
NOTE: This content is solely for purposes of information and does not substitute for diagnostic or medical advice. Talk to your doctor if you are experiencing pelvic pain, or have any other health concerns or questions of a personal medical nature.
[i] S. Yeon Youn, M. Hyung Choi, D Hwan Kim, Y. Joon Lee, H. Huisman, E. Johnson, T. Penzkofer, I Shabunin, D. Jean Winkel, P. Xing, D. Szolar, R. Grimm, H. von Busch, Y. Don, B. Lou, A. Kamen. Detection and PI-RADS Classification of Focal Lesions in Prostate MRI: Performance Comparison Between a Deep Learning-based Algorithm (DLA) and Radiologists with Various Levls of Experience. European Journal of Radiology (2021), doi: https://doi.org/10.1016/j.ejrad.2021.109894