Cross-device automated prostate cancer localization with multiparametric MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Automated cancer localization is vital for guiding biopsies, surgery, and treatment.
• Supervised learning techniques require accurate, device-specific training datasets.
• Variations in MRI devices, protocols, and field strengths create different intensity profiles, necessitating separate datasets for each, which is costly.

Purpose of the Study:

• To develop a novel method for creating cross-device and cross-protocol automated cancer localization classifiers.
• To enable the use of classifiers trained on one imaging setup for datasets acquired from different setups.

Main Methods:

• Proposed a novel method for designing classifiers that are transferable across different imaging protocols and MRI devices.
• Investigated prostate cancer localization using multiparametric MRI as an example application.
• Evaluated the efficacy of relative intensity-based methods compared to standard normalization techniques like z-score.

Main Results:

• Simple normalization techniques (e.g., z-score) were insufficient for cross-device automated cancer localization.
• The developed relative intensity-based methods successfully enabled the application of a classifier trained on one device to a test patient imaged with a different device.
• Demonstrated successful cross-device transfer of a prostate cancer localization classifier.

Conclusions:

• The proposed relative intensity-based method overcomes the limitations of standard normalization for cross-device transfer learning in automated cancer localization.
• This approach significantly reduces the cost and effort associated with creating device-specific training datasets.
• The method holds promise for improving the generalizability and efficiency of AI-driven cancer diagnostics across various clinical settings.