Object detection

A team of researchers from the NETRIX S.A. Research and Development Center has developed an innovative method for detecting objects in ultrasound imaging, based on the use of advanced deep machine learning models. The new technology achieves exceptionally high accuracy while limiting the number of measurement sensors, which makes it an attractive alternative to expensive and time-consuming classic systems. This solution opens up new possibilities in both industrial and medical diagnostics.

Experimental setup and example of measurement data: a) photo of the complete experimental setup, b) experimental tank equipped with 32 ultrasonic sensors arranged clockwise, c) example of recorded data from channel no. 1.

*Przykłady rekonstrukcji obiektów w zbiorniku testowym. Szare kółka oznaczają rzeczywiste obiekty, a czerwone – przewidywania modelu.*

The research published in the PLOS ONE journal presents an experimental system consisting of three ultrasonic transducers placed on the edge of a test tank filled with water. Objects placed inside the medium constitute the basis for the analysis of signals recorded by the system. Unlike standard solutions, the developed neural network model combines convolutional layers, responsible for the extraction of characteristic signal features, with dense layers, used to predict object parameters – such as number, position and diameter.

The achievements of the method are impressive. The classification model achieved a determination coefficient R² of 99.8%, while the regression model – 98.4%. This means an almost perfect match of the network predictions to the actual experimental data. Moreover, high accuracy was obtained using only three sensors, which significantly reduces the costs of the system and the time needed for analysis.

The developed solution demonstrates great versatility of applications. It can be used in industry, for example to detect material defects or monitor flows in tanks, as well as in medicine – to locate pathological changes in tissues. Thanks to the small number of sensors and the speed of operation, the system can be used in mobile and low-cost diagnostic devices.

The research team plans to further develop the method. In the future, the system can be extended to work in three dimensions, which will enable spatial localization of objects. It is also planned to increase the number of detected structures, which will allow for the analysis of more complex geometric and material systems.

This technology is a significant step forward in the development of ultrasound tomography, combining high accuracy with simple implementation and low costs. Its further development may revolutionize imaging both in the laboratory environment and in operational conditions.

The full publication is available in the journal PLOS ONE:
link to publication