Imaging-based classification algorithms on clinical trial data with injected tumour responses

Current microwave breast cancer imaging algorithms focus primarily on generating an image, and provide little machinery for interpretation of the image. Within-image contrast is commonly used as a performance metric, but a better reflection of the tumour detection capability of an algorithm is the difference between the maximum voxel intensities observed in images from scans of tumour-free and tumour-bearing breasts. This paper extends existing imaging algorithms by incorporating an automatic tumour detection technique that involves classification based on maximum voxel intensities. We compare results obtained from different algorithms on the data collected from healthy breast scans performed during clinical trials of a microwave radar system. We artificially inject tumour signals that are constructed based on the transmission properties of the radar system and the estimated breast tissue properties. The results provide insights into which algorithms are sufficiently robust to handle discrepancies between the real measurement data and the modeling assumptions.