Fig 1.
Video clip organised as “raw” multi–dimensional array.
Fig 2.
An overview of the proposed approach.
Ultrasound video clips are cropped to regions of interest, then processed using tensor decomposition algorithms. The resulting three components from the decomposition are fed into a classifier neural network with three LSTM layers (one per input) followed by dense layers to predict presence or absence of stenosis.
Fig 3.
Canonical Polyadic Decomposition (CPD).
An intuitive illustration of the way CPD decomposes the original “raw” data combined into a third–order tensor, resulting in a linear combination of spectral, temporal and spatial components which are independent from each other.
Fig 4.
In–depth architecture of the multi–input LSTM neural network used to predict presence or absence of stenosis.
Multi–input neural network with three inputs, the three components generated by the Tensor decomposition of the video: The frequency content, the time course, and the distribution over channels. Each input feeds into a separate LSTM layer before going through a hyperbolic tangent activation layer. The outputs are concatenated before being fed into a Dense layer followed by a ReLU activation layer, which then feeds into a single–neuron final Dense layer followed by a sigmoid activation for classification.
Table 1.
All baseline characteristics of patients whose scans were included.
Table 2.
DUS measurements stratified by site of stenosis (arterial flow, venous flow, anastomosis).
Fig 5.
Spectrograms of arteriovenous flow (whole image) with and without stenosis.
The spectrograms show the signals picked up in the whole image of arterial inflow, anastomosis, and venous outflow of fistulae with stenosis and no stenosis, which includes signals from the blood and the surrounding tissues. The Doppler readings in the vein with no stenosis are PSV (176 cm/s)–VF (475 ml/min)–diameter (2–5.8 mm) and with stenosis is diameter (<1 mm). In the anastomosis with no stenosis are PSV (386 cm/s)–diameter (4.4 mm) and with stenosis are PSV (500 cm/s)–diameter (2.4 mm). In the artery with no stenosis are PSV (187 cm/s)–VF (4950 ml/min)–diameter (5.8 mm) and with stenosis are PSV (315 cm/s)–VF (1800 ml/min)–diameter (5.2 mm).
Fig 6.
Spectrograms of blood flow only with and without stenosis.
The spectrograms show the signals picked up in the blood flow only of arterial inflow, anastomosis, and venous outflow of fistulae with stenosis and no stenosis. The Doppler readings in the arterial with no stenosis are PSV (197 cm/s)–VF (1721 ml/min)–diameter (7 mm) and with stenosis are high resistive PSV (368 cm/s)–VF (290 ml/min)–diameter (5.6mm). In the anastomosis with no stenosis are HH 07 –PSV (250 cm/s)–diameter (3.2 mm) and with stenosis are PSV (340 cm/s)–diameter (2.9 cm)–stenosis (40–50%). In the venous with no stenosis are HH 04 –PSV (100 cm/s)–VF (610 ml/min)–diameter (4–9 mm) and with stenosis are PSV (693 cm/s)–VF (490 ml/min)–stenosis (45–50%). The purple arrows show the random peaks with high intensity, which correspond to an artefact either caused by the patient movement or by the operator moving transducer out of the plane.
Fig 7.
Bootstrapped ROC curves (n = 10,000) generated from the LSTM classifier predictions along with the ROC curve for the full test set.
The non–bootstrapped AUROC was 0.821. The variation seen in the bootstrapped ROC curves shows the effects of a limited dataset size.
Fig 8.
Confusion matrix comparing the model predictions against ground truths.
The model is highly trustworthy when predicting a stenosis outcome, only mistaking one non–stenosis case for stenosis. Twenty–two stenosis cases are correctly classified as stenosis. Six stenosis cases are incorrectly classified as non–stenosis and two non–stenosis cases are incorrectly classified as stenosis.