5. Stiffness computation using DL-framework and proposed work
Stiffness measurements involve estimation of LD and IMT, state-of-the-art techniques applied in segmentation and analysis of lumen region are benchmarked in this section. Menchón-Lara et al. applied machine learning (ML) for carotid far-wall segmentation [120]. The method used multi-layer perceptron-based binary classifier to trace IMT and surmounted morphological variations in the artery. Segmentation error for LI, MA was 37.03±18.57 μm, 34.52±10.29 μm and IMT assessment error was 37.63±25.18 μm. Later, a modified algorithm by the same team applied DL-based classifier for ROI detection and was successful in almost all (99.44±0.05%) cases [2]. The estimated mean IMT error reduced to 5.8±34 µm for 67 subjects [2]. However, the study failed for near-wall analysis. Biswas et al. applied DL for lumen detection and characterisation [3]. The system used a 13-layered CNN for feature extraction (encoder), and a three-layered FCN for lumen segmentation (decoder); reported accuracy of 99.89% in LD measurement. Further, Jaccard index and Dice similarity for lumen segmentation on 407 images was 0.94 and 0.97 respectively and were effective in both far-wall and near-wall detection.
In a multi-frame approach, Tajbakhsh et al. automated frame selection, ROI localisation, and IMT measurement using DL [135]. Frames corresponding to the R wave in ECG were used as the reference in frame selection and carotid bulb (~1 cm radius) was used for ROI localisation. This improved localisation error for LI but had little impact for MA due to low-contrast and poor gradient. Moreover, the results agreed with the expert readings (p ~0.5 in ANOVA) on 44 videos from 11 subjects. Limited dataset affects the performance of DL-system so researchers used a patch-based approach. Lekadir et al. [136] used image patches on an architecture that included four CNN and three FCN layers for tissue characterization (as lipid core, fibrous and calcified). The authors applied 90,000 patches from readings of 56 subjects to train the network and were successful in 78.5% trials [136]. Perhaps, this was due to the limited dataset and quality of GT used to train the network. Studies on DL-based lumen segmentation techniques are compared in Table IV.