Adaptive Normalized Convolution for 4D reconstruction of freehand-rotated 2D TEE sequences

Major heart rhythm disorders are often treated by implanting a pacemaker, but the pacemaker leads are vulnerable and sometimes need to be replaced. Extraction of malfunctioning leads (wires) is a risky and long procedure, which is currently performed under X-ray guidance. 3D Transesophageal Echocardiography (TEE) can be used as an image guidance tool in order to reduce the radiation dose. However, due to long procedure time there is a need for thinner probes that cause less discomfort to the patient and less image blockage to X-ray. We propose a new 4D reconstruction method from 2D images acquired using a manually rotated microTEE probe. The method uses a normalized convolution approach which adapts its kernel sizes to the local density in the sampling distribution. The method is optimized and tested with 2D ultrasound sequences generated from ground truth 4D datasets with angle/phase sampling distributions from actual microTEE recordings. The adaptive method outperforms standard NC with respect to RMSE and SSIM measures for a range of distributions.