Computer Vision News - June 2023

29 Andrea Lara Motivated by the above, she introduces the first deep learning-based registration method for dynamic myocardial perfusion CT studies [ 2 ] . The proposed approach is based on a novel loss function and a recursive cascade configuration that performs 2D registration without affecting the changing contrast agent concentration, which is essential to performperfusion measurement. Moreover, this method is tested in a clinical example to quantify myocardial perfusion using an unseen patient with known aortic valve insufficiency as shown in Fig 1. Finally, she investigates other applications where temporal information is essential in the understanding of the dynamics and physiology, such as the detection and tracking the muscle tendon junction [ 3 ] . In this work, she contributed by studying different deep learning architectures that exploit the spatial and temporal information present in musculoskeletal US sequences. In the future, Andrea plans to extend her research to other dynamic imaging modalities and focus on applications that address current clinical needs in low-middle income countries (LMICs). Also, she wants to continue working to strengthen and empowering researchers from LMICs through initiatives like RISE-MICCAI and SIPAIM . Result of clinical example.: (a) CT values time curves obtained from a ROI in the LV cavity (input function u(t)) (b) Fermi-function for deconvolution (c) measured and estimated CT values-time curves in the segmented LV myocardial wall (output function y(t)). (d) difference in HU values over time for the unregistered images (misalignment of the myocardium over the sequence). (e) difference in HU values over time for the registered images (aligned LV myocardium after LCV registration). (f) calculated regional myocardial perfusion in ml/100g/min for the apical (yellow), septal (orange) and lateral wall region (red).