Neural Network-Enhanced High-Speed Atomic Force Microscopy: Eliminating Pre-Processing for Robust DNA Molecule Detection in Nanoscience

The high-speed atomic force microscope (HSAFM) has been shown to be a suitable alternative to next generation sequencing and clinical assays for the detection of genetic mutations leading to disease. This method relies on the accurate measurement of DNA molecules, which are only 2.5 nm in diameter; therefore, image quality is imperative. Current image processing algorithms, although effective, require expert supervision and lack accuracy. Due to large datasets, these algorithms are also computationally expensive. Machine learning has been shown to be an excellent identification tool that can handle large datasets while being fast and efficient once trained. This thesis proposes the use of deep learning convolutional neural networks as alternative methods to traditional image processing algorithms and object detection for use with images produced with the HSAFM. There are three distinctive neural networks included in the architecture, each with a specific task: eliminate noise frames, process the image, detect the object. This architecture can continue to be investigated, leading to more accuracy, efficiency, and robustness.