Training Universal Deep-Learning Networks for Electromagnetic Medical Imaging Using a Large Database of Randomized...

Area of Science:

• Electromagnetic medical imaging
• Deep learning applications
• Computational electromagnetics

Background:

• Deep learning excels at inverse problems in electromagnetic medical imaging.
• Current methods suffer from inaccuracies due to limited, simplified training data.
• Need for robust datasets representing complex tissue properties.

Purpose of the Study:

• Develop a novel, expansive, and diverse database for training universal deep learning models.
• Enable accurate permittivity profile reconstruction in complex electromagnetic medical imaging.
• Enhance the generalization capabilities of deep learning networks for medical imaging inverse problems.

Main Methods:

• Constructed a database of 25,000 unique objects with randomly shaped structures (ellipses, polygons) and varying electrical properties.
• Generated representative electromagnetic signals using antenna arrays irradiating and capturing scattered signals.
• Trained a custom U-net deep learning model using the generated signals for permittivity profile reconstruction.

Main Results:

• Trained U-net achieved high accuracy on diverse testing datasets.
• Structural similarity scores exceeded 0.84.
• Normalized root mean square errors were below 14%, and peak signal-to-noise ratios exceeded 33 dB.

Conclusions:

• The constructed database is practical for training generalizable deep learning networks.
• The approach enables accurate inverse problem solving in medical imaging without auxiliary algorithms.
• Demonstrates the potential for improved diagnostic accuracy through advanced deep learning techniques.