Machine learning based prediction of image quality in prostate MRI using rapid localizer images
Abdullah Al-Hayali1, Amin Komeili2, Azar Azad3
1University of Guelph, School of Engineering, Guelph Imaging AI Lab, Guelph, Ontario, Canada.
Journal of Medical Imaging (Bellingham, Wash.)
|March 4, 2024
Summary
Machine learning accurately predicts prostate MRI quality using rapid localizer sequences, identifying suboptimal exams before they occur. This approach saves time and resources by enabling early intervention for better diagnostic imaging.
Area of Science:
- Radiology
- Medical Imaging
- Artificial Intelligence
Background:
- Prostate MRI diagnostic performance relies on high-quality imaging.
- Rectal gas and distention negatively impact prostate MRI quality.
- Early detection of poor-quality MRI can prevent wasted time and resources.
Purpose of the Study:
- To develop a machine learning model for predicting prostate MRI quality.
- To assess the feasibility of using rapid localizer sequences for quality prediction.
- To enable early intervention for suboptimal MRI exams.
Main Methods:
- A dataset of 213 prostate MRI localizer images with manual rectal segmentation was used.
- Deep learning models, including 2D U-Net with ResNet-34, were employed for prediction.
- Radiomics-based classifiers were developed and compared to expert quality scores.
Main Results:
- The best model accurately predicted the quality of T2W, DWI, and ADC sequences.
- In the test set, optimal exams were achieved in over 90% for all sequences.
- A radiomics-based classifier demonstrated superior performance in quality prediction.
Conclusions:
- A radiomics-based classifier applied to localizer images accurately predicts subsequent prostate MRI quality.
- This method allows for early identification of suboptimal imaging.
- The findings support the use of machine learning for optimizing prostate MRI acquisition.
