Physiology-Informed Generative Multi-Task Network for Contrast-Free CT Perfusion

TL;DR

This study introduces MAGIC, a physiology-informed multitask GAN that translates non-contrast CT to four contrast-free cerebral perfusion maps (CBV, CBF, MTT, TTP) for stroke assessment. By integrating physics-inspired losses (including a Central Volume Principle constraint and extrema emphasis) and a Physician-In-The-Loop module, MAGIC achieves high structural fidelity and diagnostic plausibility comparable to conventional contrast-enhanced RAPID maps, demonstrated through a large UF Health dataset and a double-blind clinical evaluation. The approach promises faster, safer, and cheaper perfusion imaging with potential to expand access in resource-limited settings, while enabling high-resolution 3D perfusion maps and interactive clinician exploration. Limitations include generalizability to external datasets, capturing microvascular detail, and dependence on co-registered NCCT-CTP pairs; future work points to diffusion-based enhancements, broader cohorts, and improved temporal map encoding to strengthen clinical applicability.

Abstract

Perfusion imaging is extensively utilized to assess hemodynamic status and tissue perfusion in various organs. Computed tomography perfusion (CTP) imaging plays a key role in the early assessment and planning of stroke treatment. While CTP provides essential perfusion parameters to identify abnormal blood flow in the brain, the use of contrast agents in CTP can lead to allergic reactions and adverse side effects, along with costing USD 4.9 billion worldwide in 2022. To address these challenges, we propose a novel deep learning framework called Multitask Automated Generation of Intermodal CT perfusion maps (MAGIC). This framework combines generative artificial intelligence and physiological information to map non-contrast computed tomography (CT) imaging to multiple contrast-free CTP imaging maps. We demonstrate enhanced image fidelity by incorporating physiological characteristics into the loss terms. Our network was trained and validated using CT image data from patients referred for stroke at UF Health and demonstrated robustness to abnormalities in brain perfusion activity. A double-blinded study was conducted involving seven experienced neuroradiologists and vascular neurologists. This study validated MAGIC's visual quality and diagnostic accuracy showing favorable performance compared to clinical perfusion imaging with intravenous contrast injection. Overall, MAGIC holds great promise in revolutionizing healthcare by offering contrast-free, cost-effective, and rapid perfusion imaging.

Figures (7)

• Figure 1: Proposed MAGIC pipeline for non-contrast CT perfusion map generation. a. Neighboring NCCT slices in the z-axis are stacked to create a pseudo-RGB image and are matched with the corresponding perfusion maps to construct paired datasets. b. MAGIC consists of a generative adversarial network to co-learn between the generator and discriminator networks using novel physiology-informed loss terms. During training, the generator synthesizes perfusion maps using a UNet architecture, which are refined through adversarial learning with the PatchGAN discriminator. In the testing stage — indicated by the dashed line in the figure —the Physician-in-the-Loop (PILO) interactive module allows clinicians to visually assess and adjust the optimal ratio between non-contrast structural information and synthesized perfusion information.
• Figure 2: Sample CTP maps in doubled blinded evaluations. The CTP perfusion maps were presented in column view and given to clinicians for clinical evaluation of a patient with moderate infarction. The left image is the real (RAPID) CTP maps, and the right image is the synthetic (MAGIC) CTP maps for the same patient.
• Figure 3: A visual comparison of real CTP maps and synthetically generated maps via MAGIC shows efficacy in characterizing the brain’s hemodynamic activity from NCCT imaging alone. a. 75Y female patient presenting acute subdural hematoma layering along the right falx. Perfusion imaging shows increased TTP and increased CBV and CBF along the area of the acute subdural hemorrhage. Synthetic CTP is largely consistent with clinical CTP. b. 63Y male patient presenting patchy small vessel ischemic disease of the white matter. Perfusion imaging shows no focal perfusion defect. Synthetic CTP is consistent with clinical CTP, showing normal perfusion activity. c. 63Y male patient presenting left cerebral intraparenchymal hemorrhage. Perfusion imaging shows decreased CBF and CBV at the region of intraparenchymal hemorrhage with increased TTP throughout the left parietal, left posterior frontal, and left temporal lobes. Synthetic CTP is largely consistent with clinical CTP, showing elevated MTT and TTP with decreased CBF and CBV. d. 65Y male patient presenting encephalomalacia in the right parietal, occipital, and temporal lobes, compatible with prior infarcts. Perfusion imaging shows increased TTP with decreased CBF and CBV in the region of encephalomalacia. Synthetic CTP is consistent with trends in clinical CTP, showing elevated TTP and decreased CBV and CBF.
• Figure 4: MAGIC evaluated by SSIM and UQI separated by CTP map type and RAPID ratio. a,b. MAGIC results evaluated using SSIM and UQI metrics. MAGIC performed well in synthesizing structurally similar perfusion imaging across all map types. c, d. SSIM and UQI results stratified by the patient’s RAPID-calculated ratio of infarct core volume to tissue-at-risk volume. Among 168 test patients, 104 patients have a ratio = 0, and 64 patients have a ratio $>$ 0. c. Mean SSIM values exceed 0.790 and d. mean UQI values exceed 0.890 across all map types, indicating very high structural integrity in the synthesized CTP imaging.
• Figure 5: Illustration of the effect of adjusting the perfusion information ratio in the PILO module to emphasize encoded perfusion activity versus anatomic representation in the synthesized maps. The presented image is from a 63Y male patient presenting left cerebral intraparenchymal hemorrhage. Perfusion imaging shows decreased CBF at the region of intraparenchymal hemorrhage. Increasing perfusion encoding shows exaggeration of the decrease in CBF due to the intraparenchymal hemorrhage. By adjusting the perfusion information ratio, the physician may choose to integrate more anatomical or perfusion information in the regions of interest from the synthesized perfusion maps for radiologist-preferred, task-specific, and patient-individualized analysis of the patient’s imaging.