Abstract:
Prompt profiling of ischemic stroke is critical to provide an endovascular or pharmacological treatment, avoiding neurological repercussions in patients. Current stroke c...Show MoreMetadata
Abstract:
Prompt profiling of ischemic stroke is critical to provide an endovascular or pharmacological treatment, avoiding neurological repercussions in patients. Current stroke care protocol establishes that the first imaging study is a non-contrast computed tomography (NCCT), but its radiological findings show a poor contrast between healthy and affected tissue. Complementary, Diffusion-Weighted Imaging (DWI) allow better lesion observations, but some early lesions remain challenging to differentiate with other findings; Besides, the inclusion of FLAIR sequences, in protocol analysis, results fundamental to mismatch differences with DWI, enhancing stroke characterization, and approximating the onset of ischemic lesion. Nonetheless, the acquisition of DWI and FLAIR sequences can take extended periods, and MRI scanners have scarce availability. This work proposes a generative representation capable of synthesizing FLAIR sequences from NCCT studies, forced during training to focus on regions affected by stroke while preserving textural characteristics of ischemic lesions. The proposed approach enforces ischemic lesion information through binary and dilated class weights, preserving lesion textures into the FLAIR translation domain. In a retrospective study, 87 patients with ischemic stroke were included, with an acquisition time between modalities of less than 24 hours. The proposed method can synthesize approximations of FLAIR sequences from NCCT scans (FLAIR MS-SSIM=0.90, and lesion-bounded MS-SSIM=0.60), potentially reducing the time delay between image acquisitions in clinical scenarios with limited MRI scanner availability.
Published in: 2024 20th International Symposium on Medical Information Processing and Analysis (SIPAIM)
Date of Conference: 13-15 November 2024
Date Added to IEEE Xplore: 12 December 2024
ISBN Information: