Terra Jang

Helios Scholar

According to the CDC, cancer is the second most common cause of death. Among the types of cancers, some do not have early screening methods, increasing their mortality rates. Early general cancer screening, via blood samples from regularly scheduled check-ups, is a line of research worth delving into. This project checks for the overlap between DNA fragments circulating the bloodstream and certain cancer signaling regions, then inputs the data into a convolutional neural network (CNN) to classify as either cancer or normal. Fragments from 12 reference healthy samples were mapped out for 21 bases around their starting cytosine-guanine positions and observed for overlap with Alu repeating regions. Fragments that qualified were mapped into matrices that kept track of the fragment starts, then fed into the Convolutional Neural Network (CNN). The images were run through processes that made weighted averages of the pixels and looked for any repeated traits of the images to help with classification. The network ran the data for 50 epochs, calculating the accuracy of the classification in each epoch. The current runs of the neural network had the bin counts reduced, and the accuracy increased from a peak at 68% accuracy with 601 fragment length bins to 75% with 2 bins. With adjustments of the network parameters, the accuracy could become higher, turning cancer detection via blood sample into a viable larger-scale method.