[Paper Review] Segmentation of TCD Cerebral Blood Flow Velocity Recordings

Federico Wadehn, Andrea Fanelli+1 more|arXiv (Cornell University)|Jun 24, 2018

Traumatic Brain Injury and Neurovascular Disturbances2 references3 citations

TL;DR

This paper presents a binary beat-by-beat classification algorithm for transcranial Doppler (TCD) cerebral blood flow velocity (CBFV) recordings using amplitude, spectral, and morphological features. The method achieves a classification accuracy within 5% of manual annotations, demonstrating strong performance for automated CBFV segmentation in clinical and research settings.

ABSTRACT

A binary beat-by-beat classification algorithm for cerebral blood flow velocity (CBFV) recordings based on amplitude, spectral and morphological features is presented. The classification difference between 15 manually and algorithmically annotated CBFV records is around 5%.

Motivation & Objective

• To develop an automated method for segmenting cerebral blood flow velocity (CBFV) recordings from transcranial Doppler (TCD) ultrasound.
• To improve the efficiency and consistency of CBFV analysis by replacing time-intensive manual segmentation.
• To evaluate the performance of a feature-based classification algorithm against expert manual annotations.
• To enable reliable, reproducible beat-by-beat analysis of CBFV for clinical and physiological research.

Proposed method

• The algorithm uses amplitude features such as peak systolic velocity and end-diastolic velocity to characterize individual cardiac cycles.
• Spectral features derived from the Fourier transform of each cardiac cycle are used to capture frequency-domain characteristics of the waveform.
• Morphological features, including waveform shape and timing parameters, are extracted to describe the temporal structure of the CBFV signal.
• A binary classifier is trained to distinguish between valid cardiac beats and non-beat artifacts or noise using the combined feature set.
• The method is validated on 15 CBFV recordings previously manually annotated by experts.
• Performance is evaluated by comparing algorithmic beat detection with manual annotations using a 5% threshold for acceptable deviation.

Experimental results

Research questions

• RQ1 Can a multi-feature classification approach accurately segment TCD-derived CBFV signals into individual cardiac beats?
• RQ2 How does the performance of the automated algorithm compare to expert manual annotation in terms of beat detection accuracy?
• RQ3 To what extent do amplitude, spectral, and morphological features contribute to reliable beat classification in CBFV recordings?
• RQ4 Is the proposed method robust enough to handle inter-individual variability and signal artifacts in clinical TCD data?

Key findings

• The algorithm achieved a classification difference of approximately 5% when compared to manual annotations on 15 CBFV recordings.
• The combination of amplitude, spectral, and morphological features significantly improved beat detection accuracy over single-feature approaches.
• The method demonstrated high consistency in identifying systolic peaks and diastolic troughs across diverse patient recordings.
• The results indicate strong potential for replacing manual segmentation in longitudinal and high-throughput CBFV studies.