BACKGROUND: Cancer is one of the top killers in the United States. Though various health agencies have come up with elaborate screening strategies, they are expensive and cumbersome. Hence, they are not widely accepted. Recent study shows presence of circulating tumor cell clusters (CTCC) during early stages of Cancer. CTCC can be measured with a variety of immunohistochemistry tests, flow cytometry, and filtration tools. However, these restrict distinction of certain types of cancer, and yield from small aliquots of blood sample is low. CTCC (90 – 1000 µm) are much greater in size than normal blood cells (4 – 12 µm). HYPOTHESIS: Discrimination of such differential size by continuous scanning of a blood vessel offers an attractive unified screening tool for these grave maladies of mankind. Implementation of Machine Learning will allow for automation and improved accuracy. METHODS:For Ultrasound Detection: Yeast cells have size similar to normal blood cells (sample 1). Yeast colonies can be incubated with sugar and starch to size which mimics that of CTCC surrogate (Sample 3) and Circulating Tumor Cells surrogate (CTC) (Sample 2) by controlling time of incubation if sodium fluoride is added to arrest further growth. This made it possible to control size of the yeast colonies. Such size can be confirmed under microscope. An ultrasound phantom was made by fastening a tube in a container and pouring agar on top. The agar mimicked human soft tissue. The tube was later removed and intravenous (IV) tube connectors were added on both sides, creating a wall-less cavity, mimicking carotid artery. The samples created earlier were circulated through the phantom. Doppler Ultrasound was performed on the phantom. Key characteristics of particles of interest were identified by the study author and confirmed by mentor. Later, an IV bag and infusion pump were connected to the Ultrasound Phantom to create a closed circuit. All samples were passed through and ultrasound was performed again on the phantom, with screenshots taken every 15 seconds. For Machine Learning: Ultrasound data was processed to be input into a Convolutional Neural Network. 18 models with binary classification were tested using Google Collaboratory Program. RESULTS: Real-time Doppler signals of CTCC could be visually distinguished from normal cells, normal saline, and CTC. The most accurate model machine learning model achieved 99.98% sensitivity and 96.26% specificity in prediction of CTCC. This model was more accurate than human evaluation. This model was used to predict results for another set of data collected earlier, for which labels were not determined. The model and human reviewer concurred on 95.33% of the signals. CONCLUSIONS: This proof of principal study successfully demonstrates that CTCC surrogate detection by Doppler ultrasound is feasible. Machine learning can automate and improve this detection.
Poster Presentation at 2020 IEEE MIT Undergraduate Research and Technology Conference.