Room: Track 1
Purpose: This work describes progress in radar-based breast microwave imaging (BMI) at the University of Manitoba. A review of recent developments in phantom design, image reconstruction techniques, and applications of machine learning for tumour detection in BMI are described.
Methods: MRI-derived numerical breast phantoms were used to develop 3D-printable models. An array of 3D-printed phantoms was created, containing 66 unique phantoms. The functional form of the iterative maximum-likelihood expectation maximization algorithm used in positron emission tomography was adapted for use with literature standard radar-based microwave imaging methods, and the impact of the iterative algorithm structure on image contrast was examined. An open-access experimental dataset consisting of data from over 1250 phantom scans was developed for use with machine learning (ML) applications. The area under the curve of the receiver operating characteristic curve (AUC) was determined for both shallow and deep learning classification techniques.
Results: The application of an iterative structure to literature standard image reconstruction techniques in radar-based BMI resulted in improved image contrast compared to the non-iterative standard approaches, increasing the signal-to-mean ratio by as much as 18 dB. Deep learning approaches that make use of the structure of the measured electromagnetic data resulted in an AUC of (86 ± 1)%, while shallow learning approaches yielded an AUC of (81.3 ± 0.1)%.
Conclusion: The development of an array of MRI-derived phantoms has allowed for the evaluation of image reconstruction and ML methods on a diverse dataset. The improvements in diagnostic performance caused by the application of an iterative structure to literature standard beamformers and more sophisticated ML methods have helped to push BMI toward clinical evaluation. Further progress is required before BMI systems are suitable for clinical use, but recent progress in phantom design, image reconstruction, and ML methods have demonstrated the potential diagnostic capability of radar-based BMI.
Funding Support, Disclosures, and Conflict of Interest: The authors would like to thank the Natural Sciences and Engineering Research Council of Canada, the CancerCare Manitoba Foundation, and the University of Manitoba for providing funding for this work.