A two-step AI framework could help with quality control for prostate cancer detection at MRI, according to research published August 18 in Radiology.
The framework, which uses a deep-learning (DL) model trained on both low- and high-quality MR images, helped improve radiologist accuracy in this area, wrote a team led by Tiago Coelho from Radboud University Medical Center in Nijmegen, the Netherlands.
“When applied to prostate MRI, the framework provided a standardized method for translating image quality thresholds into almost linear gains in both AI model diagnostic performance and radiologist accuracy, resulting in a practical, performance-driven tool for quality control,” the Coelho team wrote.
Current image quality metrics do not rely on diagnostic outcomes, but rather subjective reader labels. This can affect diagnostic performance.
Coelho et al developed and evaluated a diagnostically calibrated AI framework for image quality assessment. The framework links prostate MRI quality scores to diagnostic performance.
Low-quality axial T2-weighted MRI scans with gradient-weighted class activation mapping. The low-quality MRI scans (top row) show blurred structures and artifacts making zonal and capsule boundaries indistinguishable. On the same scans with gradient-weighted class activation map overlays (bottom row), red indicates regions that were most heavily weighted by the deep learning model toward the low-quality prediction.RSNA
To develop the framework, the team trained a DL model on “the most reliable” low-quality and high-quality axial T2-weighted images. The result was a continuous image quality score. The researchers applied this score to an independent internal test set and calibrated with diagnostic performance. They focused on area under the receiver operating characteristic curve (AUC) for clinically significant prostate cancer (csPCa) detection by an AI model, and accuracy of csPCa detection by radiologists in routine clinical practice.
The single-center retrospective study analyzed nearly 12,500 consecutive prostate multiparametric MRI exams performed between 2014 and 2023 at Radboud University Medical Center. The researchers used 1,229 T2-weighted series for training and 568 multiparametric MRI exams for diagnostic testing.
They reported the following findings:
-
The DL model could differentiate between low- and high-quality images in the training set during fivefold cross-validation (AUC, 0.99)
-
For csPCa detection in the test set, the AI model AUC improved from 0.92 to 0.99 (p = 0.005) across thresholds of DL model image quality score.
-
Radiologist accuracy improved from 77% to 85% (p = 0.01).
The AI model’s ability to generalize across the full image quality spectrum “may arise from the ability of the model to detect common artifacts such as motion blur, noise, and shading,” the authors highlighted.
The authors called for future studies to include diffusion-weighted and dynamic contrast-enhanced imaging. They also suggested the framework could be applied to train image quality models across multiple centers, scanner vendors, imaging modalities, and anatomic regions.
The framework provides an objective quantitative complement to expert judgment, according to an accompanying editorial written by Tristan Barrett, MBBS, MD, from the University of Cambridge in the UK.
“If validated across centers, performance-calibrated quality assessment could redefine how radiology measures, monitors, and maintains image quality in the era of AI,” Barrett wrote.
Read the full study here.