Deep learning based assessment of hemodynamics in the coarctation of the aortaVeršnjak, Jakob (Avtor) Špiclin, Žiga (Mentor) Goubergrits, Leonid (Komentor) congenital heart diseasemagnetic resonance imagingdata synthesiscomputational fluid dynamicsblood pressure gradientmachine learningAffordable, user-friendly, non-invasive diagnostic methods, requiring minimal expert engineering knowledge, have gained prominence in recent years. This trend is also desirable for diagnosing conditions like coarctation of the aorta (CoA), a congenital heart disease characterized by aorta narrowing, leading to a significant pressure drop (PD) - a key diagnostic indicator. Invasive techniques, such as cardiac catheterization, serve as a gold standard for measuring peak-to-peak pressure gradient. However, non-invasive measurements can also be achieved through computational modeling. Utilizing numerical methods like computational fluid dynamics (CFD), patient-specific hemodynamics can be simulated based on magnetic resonance imaging (MRI). The data generation process is briefly summarized as follows. The aorta is manually segmented from 3D steady-state free-precession magnetic resonance imaging (SSFP-MRI) scans of the thoracic aorta using ZIB Amira software (version 2015.28, Zuse Institute Berlin, Germany), involving a threshold-based algorithm with manual corrections when necessary. Further, a surface model is derived from this segmentation using the vascular modeling toolkit (VMTK, Orobix Srl, Bergamo, Italy). This model represents the aorta with a discrete centerline and corresponding circular cross-sections. It is then transformed into a closed triangulated surface mesh, which serves as a boundary condition for CFD simulations. The dataset comprises images from 106 patients with CoA before treatment, 37 patients (a subset of the initial 106 CoA patients) after treatment, and 85 healthy subjects, totaling 228 cases. This cohort was also used to generate a statistical shape model (SSM), enabling the subsequent generation of synthetic cases to augment real data. Another boundary condition is acquired through 4D phase-contrast magnetic resonance imaging (PC-MRI), capturing accurate flow fields, including inlet velocity profile and peak systolic flow rates at the descending and ascending aorta, using MEVIS Flow software (version 10.3, Fraunhofer MEVIS, Bremen, Germany). In cases where PC-MRI data was unavailable, synthetic flow boundary conditions were generated using the SSM. The final dataset included the boundary conditions for a total of 228 real and 2968 synthetic cases. Finally, the hemodynamic parameters for all cases were computed solely at the peak systolic state using the commercial CFD solver Siemens STAR-CCM+ (version 13.02, Siemens PLM Software, Plano, TX, USA). These parameters include relative static pressure, wall shear stress (WSS), turbulence kinetic energy (TKE), specific kinetic energy (KE), and velocity magnitude, as they all are affected by the narrowing of the aorta. Nonetheless, the integration of CFD into clinical practice faces limitations, including computational expenses (due to computing power requirements and long simulation times) and the demand for expertise in hemodynamics simulations. To address these challenges, this thesis proposes an alternative approach utilizing deep learning (DL). Once trained, DL models provide nearly instantaneous results, avoiding the need for extensive computational resources. A comparative assessment of two DL architectures, each trained on data with a different number of dimensions is performed to predict CoA patients' hemodynamic parameters: (i) a 1-dimensional bidirectional recurrent network (1D-BRNN) and (ii) a three-dimensional convolutional neural network (3D-CNN). The 1D-BRNN is trained to predict all hemodynamics generated by the CFD along the discrete centerline points in the aorta. Conversely, the 3D-CNN can only predict the 3D static relative pressure (i.e. pressure values on the cross-sections perpendicular to the centerline) within the aorta. Furthermore, the 1D-BRNN is also subject to hyperparameter selection through cross-validation. To evaluate the accuracy of the trained models, 18 real cases, which were neither used for training nor in generating the synthetic cohort, were selected from the real cohort. The evaluation focused on the mean and standard deviation of errors (DL model's value - CFD value) for three scalar hemodynamic parameters significantly influenced by coarctation: pressure drop (PD), maximum wall shear stress (WSSmax), and maximum velocity at the stenosis region (Vmax). Furthermore, the median of the point-wise root mean squared error (RMSE) for the pressure course along the centerline is computed. Both 3D-CNN and 1D-BRNN exhibit similar accuracy and precision. The CNN is slightly more accurate in predicting pressure courses, characterized by a lower median RMSE of 3.23 mmHg, as compared to the 1D-BRNN's median RMSE of 4.25 mmHg. However, the 1D-BRNN has better precision at PD prediction, exhibiting a lower standard deviation of the error (7.03 mmHg) compared to the 8.91 mmHg evaluated for the CNN. The similarity in performance of both models suggests that compressing the 3D aortic hemodynamics into 1D-centerline representation does not result in information loss during DL model training. The results indicate that DL methods have the potential to serve as a viable alternative to CFD, particularly for PD prediction in CoA patients. According to current guidelines, an intervention is recommended if the peak-to-peak coarctation gradient (pressure drop) exceeds 20 mmHg. The model's error, with a standard deviation of 7.03 mmHg for the 1D-BRNN, would not be a limiting factor for clinical diagnosis if the confidence intervals were appropriately set.20232023-09-21 12:15:01Magistrsko delo/naloga150687VisID: 62310sl