Development of X-ray Fan Beam Coded Aperture Diffraction Imaging for Improving Breast Cancer Diagnostics

X-ray imaging technology has been used for a multitude of medical applications over the years. The typically measured X-ray transmission data, which records shape and density information by measuring the differences in X-ray attenuation throughout a material, have been used in the imaging modalities of radiography and computed tomography (CT), but there are cases where this information alone is not enough for diagnosis. In contrast, X-ray diffraction (XRD) is another X-ray measurement modality, one that typically does not produce spatially resolved 2D/3D images, but instead investigates small spatial spots for assessing material properties/molecular structures based on scattered X-rays. While XRD measurements of human breast tissue have previously suggested differences between signatures of cancerous and benign tissues, the typical diffraction system architectures do not support fast, large field of view imaging that is necessary for medical applications.In this work, an XRD imaging system was developed that can scan a 15x15 cm2 field of view in minutes with an XRD spatial resolution of 1.4 mm2 and momentum transfer (q) resolution of 0.02 Å-1. An X-ray fan beam was used to collect a 15 cm line of XRD measurements in a single snapshot, while a coded aperture is placed between imaged objects and detector, enabling XRD spectra for individual pixels along the fan beam extent to be recovered from the multiplexed measurement. Simulations were used to identify a suitable geometry for the system, while newly designed phantoms and test objects were used to evaluate the resolution/measurement quality. Upon finishing the design, construction, and characterization of the imaging system, studies on cancerous and benign tissue simulant phantoms were conducted to develop and identify top performing machine learning classification algorithms in a well-controlled study. With a shallow neural network (SNN) developed that achieved ≈99% accuracy on XRD image data, studies progressed to real human tissues. With these developments achieved, the final study was conducted where 22 human breast lumpectomy specimens were scanned and the SNN algorithm was modified for identification of human breast cancer. For 15 primary lumpectomy cases used for training and testing, an accuracy of 99.7% was achieved, with an ROC curve AUC of 0.953 and precision-recall curve AUC of 0.771. On the remaining 7 corner/rare cases present that were held out from initial training/testing (as an external dataset), an accuracy of 99.3% was achieved by the SNN, suggesting high performance along with a need for further representation of rare tissue cases in the training process to improve classifier generalization to new lumpectomy cases. This work demonstrates that fast, large field of view XRD imaging of thin samples on a millimeter spatial scale can be achieved using coded apertures. Further, the work shows that machine learning algorithms can complement this imaging modality by making great use of the multitude of input features available when each image pixel contains a full spectrum of XRD intensity vs angle values, allowing for algorithms to differentiate between cancerous and healthy tissue with higher accuracy (99.7%) compared to simple classification approaches (97.3%). Due to this promising potential, future work should seek to further the technology, by improving the spatial/spectral resolution, scan speed, and adding depth resolution, while applying the technology to useful medical tasks including (but not limited to) intraoperative surgical margin assessment, in-vivo imaging for biopsy vetting, and improved radiation therapy tumor localization.