The first step in the research will be to properly understand the problem at hand, the Elop challenges. This will be followed by a comprehensive literature review, and identification of various metrics to compare against a ground truth in image segmentation in order to generate reliable result. Elop is searching for AI techniques (machine learning, pattern recognition and deep learning) to improve the detection capability of COBRI scanner. The second step will be to investigate a detection model based on deep learning neural networks. To conduct the research, we firstly need to create a very large dataset of various synthetic cases with exactly known ground truth. Namely, a wave simulation tool such as SimSonic is used to simulate realistic cases and measure RF data for each receiver. Using Elop’s beamforming techniques, we can create synthetic images from this RF data. Such images can be validated by comparison with images from the COBRI scanner, where the synthetic and real images are based on the same concrete structure. In the next step, we use supervised learning to train a deep learning-based image segmentation algorithm to reproduce the ground truth from the synthetic images. This leads to acquiring a large dataset of various real cases with known ground truth, measured with the COBRI scanner. Finally, we take the model trained on synthetic data and use transfer learning techniques to re-train it on real data to be able to reproduce ground truth from real images. The methodologies will be published as papers.
Project leader: Kamal Raj Chapagain
Institution: ELOP AS