Upscaling Prostate Cancer MRI Images to Cell-level Resolution Using Self-supervised Learning
Keywords: MRI, Pathology, Self-supervised Learning, Image Registration
TL;DR: A novel image fusing network for upscaling prostate MRI images to cell- level resolution, leveraging registration as a key tool.
Abstract: Magnetic Resonance Imaging (MRI) plays a pivotal role in medical imaging, particularly in the diagnosis and treatment of cancers via radiography. However, one of the limitations of MRI is its low spatial resolution, which can hinder the accurate detection and characterization of cancerous lesions, especially those that are small or subtle in nature. There is a growing need for advancements in MRI technology to improve the resolution of MRI, particularly in the field of oncology, where precise detection and segmentation of tumors are crucial for effective treatment planning and optimal patient outcomes. In this paper, we proposed a self-supervised deep learning technique to upscale cancer MRI images to cell-level resolution with pathology Whole Slide Imaging (WSI). By integrating information from pathology WSIs with MRI images, this approach aims to create hybrid images that offer a more detailed and comprehensive view of cancer tissue structures. We evaluated our techniques using prostate lesions both on the similarity metrics and downstream segmentation tasks. For the similarity, our reconstructed fusion images can achieve an average 0.933 in structural similarity index. We improved lesion segmentation dice score from 57.3% to 64.0% on the test cases. Such fusion of the two imaging modalities shows promise for improving the accuracy and reliability of cancer diagnosis, guiding treatment decisions, and ultimately improving patient outcomes.
Submission Number: 24
Loading