As RF Breast Cancer Detection Research Group, we design a time-domain radar-based microwave breast screening system that operates at 2 - 4 GHz and that has been tested on patients. The system uses an antenna array to record the transmitted and reflected signals from the breast. One of the goals of our group is to develop novel machine learning techniques to detect malignancies in breast by recorded signals. As the equipment has many interacting components, being able to test proper functioning of the equipment and separating this performance measurement from the detection algorithm development process is crucial in terms of achieving a smooth data collection process during clinical trials. Therefore, we have developed a simplistic data analysis process based on controlled experiments conducted with tissue-mimicking breast phantoms. We collect healthy (i.e. baseline) and tumorous scans of same phantom multiple times with various configurations following an experimental plan. Comparing these scans in terms of energy of differences between signals recorded by same antenna pair, we hypothesize that the difference between baseline and tumour scans are significantly higher than the difference between baseline to baseline scans. For small sample sizes (in terms of number of scans), we suggest performing a simple pairwise comparison of baseline and tumour scans. As more debugging experiments performed per configuration, the number of available comparisons increases and once it reaches to a certain level we can perform more reliable statistical tests. In order to perform this statistical hypothesis testing, we suggest using a sampling with replacement schema combined with mean aggregation that avoids strong assumptions on nature of data. We present the results of our data analysis on some experiments and compare consistencies and inconsistencies of its results with an imaging algorithm.
Here is some short reports of the anaylsis we performed:
- DMAS Imaging for antenna selection
- Equipment Debugging with t-Distributed Stochastic Neighbor Embedding
- Equipment Debugging with Mean Comparison with Bootstrapping
Please see also official project group page.
