Abstract 564: Binary classification of superparamagnetic relaxometry data for cancer screening

Abstract

Abstract Introduction: Superparamagnetic relaxometry (SPMR) is an emerging technology that holds potential for use as a second-line screening modality to improve early cancer detection. During SPMR scanning, targeted superparamagnetic iron oxide nanoparticles (SPIONs) specifically bind to cancer cells and their spatial distribution can be characterized by measurement of the magnetic field relaxation following a brief excitation pulse. Highly sensitive superconducting quantum interference devices (SQUIDs) detect relaxation of clusters of SPIONs bound to small tumors. Challenges inherent to the SPMR technology include measurement noise, as well as the competing influence of SPION uptake by healthy organs (namely the liver), which also contributes to the overall SPMR signal. Hence, manual and stand-alone classification of the SPMR data into positive (i.e., the subject has cancer) or negative (i.e., the subject does not have cancer) screen results can be erroneous. Methods: We employed a data-driven approach based on Gaussian process (GP) formulation tailored to SPMR datasets to systematically quantify the probability of cancer. In silico, we simulated the SPION uptake process and generated SPMR signals that closely resembled experimental data collected in mouse models of cancer. We investigated the classification accuracy for different amounts of SPION accumulation within the tumor, as well as different levels of measurement noise (coefficient of variation (CV)). In a phantom study, a mouse liver was simulated by clustering together nine cotton swabs containing a total of 150 μg of immobilized SPIONs, while a mouse tumor was simulated by a single cotton swab containing either 9.4 μg or 14.4 μg of immobilized SPIONs. An additional nine cotton swabs containing 32.3 μg of immobilized SPIONs (<5 μg per phantom) were evenly distributed within the scan plane to represent background SPIONs not bound to the tumor or liver. For each of the tumor phantoms, 18 datasets were collected using a magnetic relaxometry device (Senior Scientific LLC) by moving the phantom to 18 different locations. Moreover, 10 datasets were collected without using the tumor phantom to represent the expected signal from healthy mice. In each iteration, the background SPION phantoms were randomly relocated within the scan plane. Results: Our in silico analysis for tumor accumulations of 3% and 5% of the injected SPION dose achieved 87% and 97% classification accuracies, respectively, when CV=0 and 75% and 93% when CV=0.015. Similarly, in our phantom study, classification accuracies of 87.5% and 96.4%, respectively, were reported for the 9.4 μg and 14.4 μg tumor phantoms. Conclusion: Using a data-driven GP model, tumor-status classification accuracies of up to 96.4% were achieved in SPMR phantom datasets. In the future, we plan to evaluate the accuracy of our classifier in preclinical settings using animal datasets. Citation Format: Javad Sovizi, Sara L. Thrower, David Fuentes, Wolfgang Stefan, John D. Hazle, Kelsey Mathieu. Binary classification of superparamagnetic relaxometry data for cancer screening [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 564. doi:10.1158/1538-7445.AM2017-564