(Sunday, 7/14/2019) 4:00 PM - 5:00 PM
Room: Exhibit Hall | Forum 1
Purpose: To assess the performance of a neural network classification algorithm on high-resolution X-ray diffraction scans of breast tissue and compare with previously-validated linear methods. The results of this experiment will help to determine what classification method to use as we continue construction of an ex-vivo margin assessment tool for intraoperative use.
Methods: An X-ray diffraction imaging system developed by us was used to image multiple breast lumpectomy specimens, post-formalin-fixation, via pencil beam raster scanning. This provided spot-specific diffraction signatures for each specimen and served as the characteristic data for training and classification. We constructed a neural network using previously-acquired data and tested it against our existing cross-correlation method in classifying a newly acquired specimen. The resultant classification map for each classifier was overlaid with an image of the sample and compared with histopathological analysis of the specimen to evaluate overall performance. The neural network was constructed by using 70% of data to train, 15% to validate, and 15% to test. Subsequent cross-validation with logistic regression was also performed.
Results: Our neural network showed approximately 10% greater accuracy, 20% higher true positive rate, and only 2% higher false positive rate over the cross-correlation method when compared to histopathological results for the new specimen. These results indicate a clear advantage in the neural network’s non-linear approach.
Conclusion: This work demonstrates a potentially significant improvement that can be conferred to a high-resolution intraoperative margin assessment system by classifying diffraction data via a properly constructed neural network.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
