All research

Physiological signal processing · 2023

Deep learning for cerebral blood-flow analysis using transcranial Doppler

Deep learning methods that restore transcranial Doppler flow-velocity waveforms and distinguish healthy volunteers from patients in neurocritical care.

QSCare · Qatar University High Impact Grant QUHI-CENG-22/23-548

Method overview. Transcranial Doppler ultrasound produces a spectrogram, a flow-velocity algorithm extracts the maximal velocity waveform, the waveforms are cleaned and restored with a CycleGAN, and a two-stream DopplerNet2+ classifier labels each recording as healthy or intensive care.
Figure 2 from Kanti Podder et al. (2023). From Doppler spectrograms to maximal flow-velocity waveforms, CycleGAN restoration and two-stream classification.View full-size figureSource article© 2023 Elsevier Ltd · author reuse

Approach

Can middle cerebral artery flow-velocity waveforms distinguish healthy volunteers from patients in neurocritical care?

The brain depends on a constant blood supply, and transcranial Doppler can monitor flow velocity in its major arteries without radiation or surgery. Signal quality depends heavily on the operator, so recordings often contain noise and motion artefacts. This project extracted maximum flow-velocity waveforms from the middle cerebral artery, restored low-quality segments with a 1D-CycleGAN, and developed Self-ONN classifiers for healthy and intensive-care recordings.

  1. Extract flow-velocity waveforms
  2. Clean and restore segments
  3. Classify healthy and ICU signals

Restoring low-quality flow-velocity waveforms

Biomedical Signal Processing and Control · 2023

The study used a public dataset recorded over the middle cerebral artery with a portable ultrasound system and a 1.75 MHz transducer. It contains 16 recordings from 6 healthy volunteers and 29 recordings from 12 patients in neurocritical care with hydrocephalus, traumatic brain injury, or intraparenchymal or subarachnoid haemorrhage. The maximum flow-velocity waveform was extracted from each Doppler spectrogram and split into 1,024-sample segments.

Trained annotators labelled each segment by signal quality. A 1D-CycleGAN with ResNet generators and PatchGAN discriminators then learned to map low-quality segments to high-quality waveforms. It reduced baseline drift and distortion in low-quality signals, while high-quality signals changed very little.

Participants
18
Segments kept after cleaning
3,243
Low-quality training segments after 75% overlap
1,760
Signal-quality labels from Figure 4 of Kanti Podder et al. (2023).
SegmentsUse
High quality2,803Kept
Low quality440Kept and restored
Mediocre2,061Removed
Corrupted164Removed
Four flow-velocity waveform plots. The high-quality signal has regular pulses, the low-quality signal is distorted but recognisable, the mediocre signal contains abrupt drops, and the corrupted signal has no clear pulse pattern.
Figure 3 from Kanti Podder et al. (2023). Examples of high-quality, low-quality, mediocre and corrupted signals.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
CycleGAN diagram. One generator converts bad-quality waveforms to good quality and a second converts good-quality waveforms to bad quality. Two discriminators, cycle-consistency losses and identity losses link the two directions.
Figure 5 from Kanti Podder et al. (2023). The 1D-CycleGAN used to restore middle cerebral artery waveforms.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
Four plots of middle cerebral artery waveforms, two from healthy and two from intensive-care recordings. Each overlays the original waveform in blue and the CycleGAN-processed waveform in orange.
Figure 7 from Kanti Podder et al. (2023). Original and CycleGAN-processed waveforms for the healthy (a, b) and intensive-care (c, d) classes.View full-size figureSource article© 2023 Elsevier Ltd · author reuse

Two-stream classification with DopplerNet2+

Biomedical Signal Processing and Control · 2023

DopplerNet is a one-dimensional Self-ONN classifier built from inverted residual blocks and pool-wise dense connections. DopplerNet2+ extends it with two parallel streams, one for the original waveform and one for the CycleGAN-processed waveform, and combines their features before classification.

Models were tested with five-fold cross-validation separated by recording session. The two-stream model gave the best result, correctly identifying 98.86% of healthy segments and 99.96% of intensive-care segments.

Accuracy
99.45%
Specificity
99.37%
Area under the ROC curve
1.00
Best result for each input and model design, from Table 4 of Kanti Podder et al. (2023).
ParametersAccuracyF1 scoreSpecificity
Original waveform199,66697.71%97.71%97.61%
CycleGAN-processed waveform587,37895.55%95.54%95.36%
Both as two input channels587,66696.64%96.64%96.43%
DopplerNet2+ with two streams1,174,75499.45%99.45%99.37%
DopplerNet2+ architecture. The original waveform and the CycleGAN-processed waveform pass through separate streams of Self-ONN blocks. Features from both streams are concatenated and sent to a classifier that outputs healthy or ICU.
Figure 9 from Kanti Podder et al. (2023). The two-stream DopplerNet2+ architecture.View full-size figureSource article© 2023 Elsevier Ltd · author reuse
Confusion matrix with 2,432 healthy and 2,814 intensive-care segments classified correctly, 28 healthy segments misclassified and 1 intensive-care segment misclassified, beside a ROC curve reaching an area of 1.00.
Figure 12 from Kanti Podder et al. (2023). Confusion matrix (a) and ROC curves (b) for DopplerNet2+.View full-size figureSource article© 2023 Elsevier Ltd · author reuse

Attention-based classification of clear waveforms

Diagnostics · 2023

The second study classified clear waveform segments directly, without generative restoration. Segments were annotated as clear or corrupted, leaving 1,501 healthy and 1,742 intensive-care segments of 1,024 samples at 217 Hz.

Self-ResNet18 replaces the convolutional layers of ResNet18 with Self-ONN layers, and Self-ResAttentioNet18 adds four multi-head attention layers. Each design was tested at three Self-ONN orders using five-fold validation separated by recording session. Self-ResAttentioNet18 at order 1 performed best, correctly identifying 96.63% of healthy segments and 95.51% of intensive-care segments.

Accuracy
96.05%
Specificity
96.09%
Area under the ROC curve
0.99
Model comparison from Table 4 of Nisha et al. (2023). Q1, Q3 and Q5 are the Self-ONN orders.
AccuracyPrecisionF1 scoreSpecificity
Self-ResNet18 Q193.46%93.48%93.46%93.39%
Self-ResNet18 Q394.91%94.92%94.91%94.88%
Self-ResNet18 Q593.31%93.32%93.30%93.26%
Self-ResAttentioNet18 Q196.05%96.06%96.06%96.09%
Self-ResAttentioNet18 Q395.31%95.34%95.31%95.37%
Self-ResAttentioNet18 Q592.97%93.01%92.97%93.03%
Self-ResAttentioNet18 architecture. A one-dimensional input passes through Self-ONN residual blocks with four multi-head attention layers, followed by flattening, a fully connected layer and a log-softmax output.
Figure 6 from Nisha et al. (2023). The Self-ResAttentioNet18 architecture.View full-size figureSource articleCC BY 4.0
Confusion matrix with 1,519 healthy and 1,596 intensive-care segments classified correctly, 53 healthy and 75 intensive-care segments misclassified, beside ROC curves with an area of 0.99.
Figure 9 from Nisha et al. (2023). Confusion matrix (a) and ROC curves (b) for Self-ResAttentioNet18 at order 1.View full-size figureSource articleCC BY 4.0

My contribution

Across the two studies, I worked on waveform extraction and CycleGAN restoration, classifier development, data curation and preprocessing, MATLAB software, analysis, visualisation, and manuscript writing.

Study scope

Both studies used one public dataset from 6 healthy volunteers and 12 patients, with testing separated by recording session. The patient class combined several neurocritical conditions, and signal-quality labels were assigned manually. Larger and more diverse cohorts, condition-specific labels and automated quality assessment are needed before clinical use.