Click here to

Session: Radiobiology and Small Animal Systems [Return to Session]

Ex Vivo Quantitative X-Ray Fluorescence Analysis of Gold Nanoparticle Based On Machine Learning

D Hara*, W Tao, N Dogan, A Pollack, J Ford, J Shi, University of Miami, Miami, FL


TU-D930-IePD-F3-1 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 3

Purpose: X-ray fluorescence Imaging attracts high attention in nanomedicine research because of its ability to accurately localize and quantify deep-seated metal nanoparticles in vivo. However, XFI suffers from long imaging times and high radiation dose because ample radiation is necessary to capture enough x-ray fluorescence signal (XRFS) to contrast from Compton scattering background noise with current standards in XRFS processing. Therefore, we describe the use of machine learning for signal processing to model complex relationships in the acquired energy spectrum from x-ray fluorescence signals to improve gold nanoparticle (GNP) quantification sensitivity and reduce signal-to-noise ratio.

Methods: 15nm PEGylated GNPs were conjugated to anti-PSMA antibodies and injected intravenously through the tail vein to nude mice bearing subcutaneous LNCaP prostate cancer xenografts. 24hrs after injection, various tissues (brain, heart, lung, muscle, pancreas, skin, spleen, liver, kidney, blood, and tumor) were excised and fixed in paraformaldehyde. GNPs were quantified by measuring L-shell XRFS on an in-house developed dual-modality XFCT/transmission CT imaging system. Ground truth GNP concentrations in tissue were confirmed with inductively coupled plasma mass spectrometry (ICP-MS). Deep learning architecture was created with Tensorflow. The multilayer neural network was developed using two hidden layers each with 64 neurons over 500 epochs. An adam optimizer with mean squared error loss function was employed for training a regression-based machine learning model using a data set of 500 XRFSs labeled with ICP-MS measurements. A validation set of 100 signals from ex-vivo tumor tissue were used.

Results: The multilayer neural network predicted GNP concentrations in tumors between 0µg/mL to 300µg/mL concentrations with an R2 of 0.84. An uncertainty of less than three percent mean absolute error was achieved.

Conclusion: This study validated the effectiveness of our machine learning-based method for ex vivo XRF GNP quantification analysis, assisting in exploring mechanism of therapeutic GNPs in oncology studies.


K X-ray Fluorescence (KXRF), Signal Processing, Quantitative Imaging


IM- Other (General): Nanoparticles (general)

Contact Email