Addressing Acquisition Variability in DWI on Prostate Cancer with Unsupervised Methods

Abstract

Diffusion-weighted MR images are typically obtained as multiple acquisitions with multiple diffusion-sensitizing gradient directions. Due to molecular motion, some acquisitions suffer from signal loss at random locations. This affects cancer conspicuity and degrades the diagnostic efficacy of DWI. We propose an agglomerative clustering-based unsupervised method to address this. The model automatically rejects acquisitions of voxels that are likely to be corrupted by bulk motion and lack coherence with the rest of the acquisitions. We observed that this method both reduces the DWI signal variability and enhances the cancer detection accuracy.