Room: Exhibit Hall | Forum 8
Purpose: Previously we demonstrated the ability to identify cancer in resected tumors using an X-ray diffraction approach. The approach utilized an X-ray diffraction imager along with a correlation-based classifier for identifying cancer. Here we describe a viable, high-resolution intraoperative cancer identification system that combines X-ray diffraction imaging with a machine learning tissue classifier to achieve high-resolution, high-accuracy cancer imaging.
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. Each specimen was scanned using an optimized beam (with reduced keV and mAs to permit high-throughput scanning) at 1mm resolution. A diffraction signature was extracted for each voxel in the image and analyzed using SVM and random forest classifiers to generate spot classifications of "healthy", "fibroglandular", or "cancer". The resultant classification map was overlayed with the image of the sample and compared with the histopathologic analysis of the specimen to evaluate accuracy.
Results: The classified image showed a clear region corresponding to the tumor, which closely matched the discernible cancer margin identified through histopathologic analysis. The optimized scan settings showed complete agreement in classification of adipose and cancerous tissue, and an average of 82% agreement in classifying fibroglandular tissues compared to the previous high-energy setting. Due to decreased anode heating and lower exposure time of the optimized settings, the scan was performed in less than half the time of the previous scanner and with 1 mm resolution. Despite limited training data, the machine-learning classifiers showed reduced error compared to the original cross-correlation classifier.
Conclusion: This work shows the viability of a high-resolution intraoperative cancer detection imaging system capable of imaging tumor margins with mm resolution using X-ray diffraction combined with machine-learning methods.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the NC Biotech Center Biotechnology Innovation Grant.