Cross-device automated prostate cancer localization with multiparametric MRI.

Area of Science:

• Medical Imaging
• Machine Learning in Oncology
• Radiomics

Background:

• Supervised classification for prostate cancer localization requires accurate training data.
• Variations in MRI device protocols and field strengths create different intensity profiles, necessitating device-specific datasets.
• Adapting existing classifiers across different devices/protocols is highly desirable to reduce data acquisition costs.

Purpose of the Study:

• To develop a novel method for adapting supervised classifiers trained on one MRI device/protocol for use on data from another.
• To enable cross-device automated prostate cancer localization using multiparametric MRI.
• To directly utilize T2-weighted MRI images without ad hoc normalization.

Main Methods:

• A novel relative intensity-based method was developed for cross-device classifier adaptation.
• The method was tested for prostate cancer localization using multiparametric MRI data.
• Data were acquired from 18 biopsy-confirmed cancer patients using two different 1.5-T MRI scanners (GE Excite HD and Philips Achieva).

Main Results:

• Simple normalization techniques like z-score are insufficient for cross-device automated cancer localization.
• The proposed relative intensity method successfully enabled the use of a classifier trained on one device for a test patient imaged with another device.
• The method allows direct use of T2-weighted MRI images, eliminating the need for ad hoc normalization.

Conclusions:

• The developed method facilitates cross-device automated classification for prostate cancer localization.
• The approach effectively utilizes T2-weighted MRI images without requiring subject-specific normalization.
• This technique offers a cost-effective solution for adapting machine learning models in medical imaging across different hardware.