European Radiology: AI-assisted biparametric MRI prostate cancer surveillance

Prostate cancer (PCa) is the second most common malignancy in men worldwide, accounting for more than
one in five cancer diagnoses in men.
In 2020, there were an estimated 1.
4 million new cases and 375 000 PCa-related deaths
worldwide.
Clinically, prostate-specific antigen (PSA) screening introduced to detect PCa has led to overdiagnosis
of low-risk PCa.
Active surveillance (AS) is a widely used treatment plan that reduces overtreatment
of low-risk PCa.
Men who receive active surveillance are closely monitored with PSA measurements and repeat biopsies to detect disease progression early and maximize
subsequent treatment outcomes.

Serial MRI provides a noninvasive method for monitoring patients with PCa, which radiologists often use to compare different patients' condition at presentation and report on radiological stability or progress
.
The PRECISE guideline is currently the best practice for reporting continuous MRI results in the AS setting, based on changes in the size and characteristics of
PCa lesions.
However, the diagnostic accuracy currently reported is still too low to support MRI-based monitoring as a reliable alternative
to repeat biopsies in AS protocols.

With the development of technology, artificial intelligence (AI) can improve MRI monitoring
.
Many studies have applied deep learning (DL) to automate and characterize
PCa on MRI at a single time point.
However, to date, none of the studies have explored the use of differentiated information
in serial MRI of the prostate.

Recently, a study published in the journal European Radiology evaluated the feasibility of artificial intelligence-assisted biparametric MRI (bpMRI) in automatically detecting csPCa, providing technical support
for faster and more accurate clinical diagnosis of PCa.

This retrospective study included 1513 patients who underwent bpMRI (T2+DWI) between 2014 and 2020, of whom 73 received at least two consecutive bpMRI scans and repeat biopsies
.
In this study, a deep learning PCa detection model was developed to generate heat maps
of all PIRADS≥2 lesions.
Use heatmaps of each patient's previous and current examinations to extract different volume and likelihood characteristics to reflect interpretable changes
between tests.
Train a machine learning classifier and predict csPCa (ISUP>1)
in the current examination from these features based on the biopsy results.
Cross-validated diagnostic accuracy was compared using ROC analysis
.
The diagnostic performance of the best model was compared
with the radiologist score.

The model of the current study (AUC 0.
81, CI: 0.
69, 0.
91) had higher diagnostic accuracy than the model of the current study alone (AUC 0.
73, CI: 0.
61, 0.
84) (P = 0.
04).

The addition of clinical variables further improved diagnostic performance (AUC 0.
86, CI: 0.
77, 0.
93).

The diagnostic performance of the AI model was significantly better than that of radiologists (AUC 0.
69, CI: 0.
54, 0.
81) (P = 0.
02).

Pictured is a 70-year-old man with a central gland zone lesion (delineated by the purple outline in each image) that AI correctly diagnoses as non-csPCa.

While the initial biopsy result on the previous scan was negative for PCa, the targeted repeat biopsy at this test showed non-csPCa (Gleason 3 + 3 = 6).

The T2 weighted images (a, b) and ADC plots (c, d) are shown, along with the corresponding detection heatmaps (e, f)
for previous and current inspections.
The detection heatmap of the second examination showed little change in volume and likelihood score
.
The radiologist gave PIRADS a score of 4

This study shows that AI-assisted prostate MRI monitoring proposed in this study can provide clinical changes and information related to diagnosis, and has good diagnostic accuracy
.

Original source:

C Roest,T C Kwee,A Saha,et al.
AI-assisted biparametric MRI surveillance of prostate cancer: feasibility study.
DOI:10.
1007/s00330-022-09032-7