Discussion
From this, several conclusions can be derived. Important studies in
stiffness assessment reported ambiguity caused by the non-uniformity in
external distance measurement. Further, the external distance
measurement does not consider the internal contour of the artery which
is important. Researchers presented a difference in reference values of
PWV owing to local or regional measurement. The image-based techniques
propose an added advantage of presenting a clear idea of
atherosclerosis. These techniques are either fully automated or semi,
based on supervisor involvement. Moreover, these methods required
arteries to be straight and horizontal for best performance. Further,
noise and artefacts deteriorated the assessment. Evidence presented here
is conclusive on the success of DL in lumen characterisation and
segmentation. This paper proposes deep learning-based stiffness
measurement from cine-loop, a short (ultrasound) video, of the carotid
artery. Cine-loop capture IMT and LD variations over many cardiac cycles
and averaging over multiple frames diminish the effects of noise.
Recently, Patel et al. [137] estimated elasticity using ELM. ELM was
applied for ROI localisation, and in 100 images maximum-IMT and
maximum-LD error were 20 mm and 91 mm, respectively. However, the study
estimated Young’s modulus and at present little is known about the
correlation between modulus of elasticity and cfPWV in a clinical setup.
The proposed method includes four stages of operation (illustration in
Figure 4.): pre-processing, ROI localisation: extraction of lumen
region, LD/IMT assessment and stiffness computation. The purpose of
pre-processing is auto-crop of textual data with emphasis on tissue
region. Pre-processing involves tasks such as brightness transformation,
geometric transformation edge detection and image restoration in
preparing the ultrasound image for the training and testing. Manual
tracing by experts is used as GT for training. Networks for lumen and/or
intima-adventitia are trained using separate GT. Further, down sampling
is applied to improve processing efficiency. Often ECG recordings are
made along with the carotid scan. R-wave is used to synchronize frames
in consecutive cardiac cycles. IMT is processed from frames 5 seconds
after R-wave to capture the extreme values during the two phases of the
cardiac cycle: systole and diastole. Lumen region and IMC region are the
ROI in stiffness computation leading to stiffness computation as
analysed in Section 2 of this paper. ROI localisation is applied twice
in the entire process: one for the identification of LI-far region and
two for recognition of LI-near zone. LI-far facilitates the computation
of IMT. Both LI-far and LI-near were required for the assessment of LD.
LD/IMT assessment was performed using a poly-line distance method.