[home] [Personal Program] [Help]
tag
16:00
0 mins
Multi-perspective photoacoustic imaging compounding based on a deep learning approach
Amir Gholampour, Navchetan Awasthi, Jan-Willem Muller, Hans-Martin Schwab, Min Wu, Josien Pluim, Richard Lopata
Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00



Amir Gholampour (Eindhoven University of Technology)
Navchetan Awasthi (Eindhoven University of Technology)
Jan-Willem Muller (Eindhoven University of Technology)
Hans-Martin Schwab (Eindhoven University of Technology)
Min Wu (Eindhoven University of Technology)
Josien Pluim (Eindhoven University of Technology)
Richard Lopata (Eindhoven University of Technology)


Abstract:
Photoacoustic imaging has great potential in providing information about tissue morphology. Linear or curved arrays are commonly used to perform this imaging technique. However, the resolution, contrast, and field of view are limited. Multi-perspective photoacoustic imaging (MP-PAI) can overcome these limitations by increasing the area for receiving photoacoustic signals using multiple transducers. The conventional approach to fuse the images from individual transducers is to average, which is simple and straightforward, however, this is not always optimal. In this work, we propose to employ a deep neural network using a modified U-net architecture to improve the image fusion step (compounding). To generate the dataset, a simulation framework is implemented which consists of a Monte-Carlo step to simulate the optical fluence and the k-Wave toolbox to simulate acoustic wave propagation. In the simulated phantoms, the optical and acoustic properties are randomized. The positions and orientations of the transducers and the light sources are randomized as well. The dataset contains 1400 set of images that are separated into 1000 for training, 200 for validation, and 200 for testing. The compounded images, in general, resemble the initial pressures that were used for the simulations, revealing great improvement in terms of resolution and contrast. To compare the results of both conventional and deep learning-based compounding approaches against the initial pressure, SSIM and PSNR evaluations are used. The average SSIM improved from 0.08 in conventional compounding to 0.84 in the deep learning approach, and the average PSNR improved from 5 dB to 30.14 dB. Overall, the preliminary results are promising and demonstrate the feasibility of this approach for employment in the experimental framework. In future studies, the training dataset can be extended by increasing complexity of the phantoms to improve the robustness of this method.