15:20
Heart  II
Chair: Theo van Walsum
15:20
15 mins

Estimation of cardiac tissue conductivity using confirmatory factor analysis
Miao Sun, Natasja M.S. de Groot, Richard C. Hendriks
Abstract: The electrical conductivity of cardiac tissue plays an important role in the development of heart rhythm disorders. Areas of impaired electrical conductivity may be the arrhythmogenic substrate that causes heart rhythm disorders. In this work, we present an effective method to estimate the conductivity of atrial tissue from epicardial unipolar electrograms using a highresolution electrode array [1].
These arrays measure the propagation of the extracellular potential in the cardiac tissue at multiple positions simultaneously. Given this data, it is in principle possible to estimate the conductivity for groups of cells by solving an inverse problem. Prior to estimating the conductivity, several other parameters in the data model need to be determined, among which the ionic currents of the cells. Recently, it was proposed in [2] to do this based on an impulse response model. However, this method does not explicitly make use of the spatial structure of the multielectrode data. In the current work, we exploit the cross power spectral density matrix (CPSDM) model of the electrograms to estimate the conductivity. This enables us to utilize the spatial correlations among the measurements to estimate the conductivity parameters.
Estimating the conductivity parameters from the measurements is an illposed problem due to the large number of unknown parameters in the data model. To estimate the parameters in the CPSDM model, we need to verify whether the solution for estimating the target unknowns can be obtained with the given amount of data. In this work, the confirmatory factor analysis (CFA) method is applied to solve this problem. By analyzing the identifiability conditions in the CFA problem, we can find the relationship between the desired resolution and the required amount of data. After the model is identified, the target parameters are estimated by solving a constrained optimization problem. Based on the fact that the conductivity parameters are shared among multiple frequencies, we further proposed the simultaneous CFA across multiple temporal frequencies to estimate the tissue conductivity. This helps to easier satisfy the identifiability conditions. Experimental results on the simulated data and the clinical data demonstrate the robustness of the proposed method for conductivity estimation.
[1] A. Yaksh, L. J. van der Does, C. Kik, P. Knops, F. B. Oei, P. C. van de Woestijne, J. A. Bekkers, A. J. Bogers, M. A. Allessie, and N. M. de Groot, “A novel intraoperative, high resolution atrial mapping approach,” Journal of Interventional Cardiac Electrophysiology, vol. 44, no. 3, pp. 221–225, 2015.
[2] B. Abdi, R. C. Hendriks, A.J. van der Veen, and N. M. de Groot, “A compact matrix model for atrial electrograms for tissue conductivity estimation,” Computers in biology and medicine, vol. 107, pp. 284–291, 2019.

15:35
15 mins

Qualitative evaluation of metrics used in deep learning for medical imaging segmentation
Leonardus van den Oever, W.A. van Veldhuizen, Geertuida de Bock, Peter van Ooijen
Abstract: Introduction:
Organsatrisk contouring is time consuming and labour intensive. Automation by deep learning algorithms would decrease the workload of radiotherapists and technicians considerably. However, the amount of metrics used for the evaluation of deep learning algorithms make the results of many papers difficult to interpret. In this paper, a qualitative evaluation is done on two metrics to assess whether their values indicate clinical usability.
Method:
377 CT volumes with heart delineations were randomly selected and divided into a training subset (303 volumes), a tuning subset and an internal validation subset (both 37 volumes). Three Unet architectures with dilated convolutional layers, one each for axial, coronal and sagittal planes, were trained on the optimal ratio between the number of slices with and without heart. The predictions of the three neural networks were combined by majority vote.
120 CT slices with the final prediction were shown to two radiologists, who were then asked to examine each slice independently whether they would accept the prediction, when making contours themselves in the clinic, and if there were small mistakes. For each slice, the scores of this qualitative evaluation were then compared with an overlapbased metric, the SørensenDice coefficient (DC), and an edgebased metric, the Hausdorff distance (HD).
Results:
A ratio of slices without heart versus ones with heart of 1:5 resulted in the best performance, reaching a mean DC of 0.91 on the internal validation dataset. The mean DC and HD for the accepted predictions were 0.91±0.07 and 10.03±7.68, while for the rejected predictions, they were 0.86±0.07 and 18.67±9.96 respectively. The mean of the DC rose 0.03% in predictions that were just accepted compared to ones that were just rejected. For the mean HD, the drop between the two types of predictions was 26%.
Conclusion:
We show that an edgebased metric gives better agreement with the assessment of a radiologist than an overlapbased metric, due to the decline of the HD to be larger than the rise of the DC between rejected and accepted predictions.

15:50
15 mins

High resolution plane wave compounding via a model based neural architecture
Nishith Chennakeshava, Ben Luijten, Oded Drori, Massimo Mischi, Yonina C. Eldar, Ruud J. G. van Sloun
Abstract: Ultrafast ultrasound imaging has enabled a plethora of new ultrasound imaging applications, but primarily relies on coherent Plane Wave (PW) compounding to obtain sufficient spatial resolution, and contrast. While the acquisition of multiple PWs for every image frame boosts spatial resolution, and contrast, this incurs a loss in temporal resolution. To overcome this trade off, we propose a model based deep learning strategy for High Resolution (HR) PW compounding from a limited number of PW transmits.

16:05
15 mins

Joint learnning of modelbased beamforming with sparse channel arrays
Ben Luijten, Iris A.M. Huijben, Frederik J. de Bruijn, Harold A.W. Schmeitz, Massimo Mischi, Ruud van Sloun
Abstract: Ultrafast ultrasound (US) imaging has gained popularity thanks to its high frame rate, however it suffers from high data bandwidth requirements. To reduce data rates, the active receive elements could be pruned, yielding sparse arrays, which are subject to recent research. Here, instead of manual design, we adopt Deep Probabilistic Subsampling (DPS) to learn an optimal set of receive channels, combined with adaptive beamforming by deep learning (ABLE). We allow this model to lean a different sparse array per plane wave angle, and jointly train it with ABLE to optimize image quality.

