[home] [Personal Program] [Help]

---

Session: Vascular I
Session starts: Thursday 26 January, 14:30
Presentation starts: 15:15
Room: Room 559

---

Camilo Cano-Barrera (Eindhoven University of Technology)
Nastaran Mohammadian Rad (Maastricht University)
Marc van Sambeek (Catherina Ziekenhuis Eindhoven)
Richard Lopata (Eindhoven University of Technology)
Min Wu (Eindhoven University of Technology)

Abstract:
Multispectral photoacoustic imaging (sPAI) is an emerging modality that allows real-time, non-invasive, and radiation-free measurements of tissue, benefitting from their optical contrast. sPAI is ideal for morphology assessment in arterial plaques, where plaque composition provides relevant information on the progression of the plaques and their vulnerability. However, since sPAI is affected by spectral coloring, general spectroscopy unmixing techniques don't provide reliable identification of the sample composition. In this study, we demonstrate the capabilities of deep learning (DL) techniques for the classification of plaque composition from sPAI. We performed ex-vivo sPAI on carotid endarterectomy samples (n=9) in a water tank at multiple acquisition angles (around 180 degrees with a step of 30 degrees). Images were acquired at wavelengths from 500 nm to 1300 nm (in steps of 5 nm) to capture the main spectral features of the different chromophores in the plaques. We transformed the raw sPAI modulated signal into a feature space. The features vector was used as input to train a two-layer convolutional neural network (CNN) architecture to label distinct components in human carotid plaque samples. Histology is used to segment regions of interest within the plaques and label the spectra of plaque materials. The test set is compared with the histology and the spectral unmixing results using an established blind unmixing method. Results show the feasibility of automatic multi-label image segmentation of sPAI using deep learning to differentiate constituent regions within a plaque. The plaque compositions such as lipids, collagen, smooth muscle cells, and hemorrhages can be well identified, and the decomposition results generally agree with the corresponding histological staining. In regions with low amplitude signals and clutter, the network outperforms state-of-the-art blind unmixing techniques. Moreover, with the proposed DL-based classification approach, there is no need to perform an extensive fluence correction to identify multiple materials, which is usually required in most unmixing techniques. All the results demonstrate the great potential to characterize plaque composition with the DL-based unmixing method. In future work, we will expand the training set by scanning plaques with surrounding tissue and exploring U-net networks to exploit the spatial information of the samples.