Graduate and Postdoctoral Studies
Integrating Coherent Anti-Stokes Raman Scattering Imaging and Deep Learning Analytics for High Precision, Label Free Cancer Diagnosis
Wednesday, August 2, 2017
to 4:00 PM
200 Brockman Hall for Physics
Coherent anti-Stokes Raman scattering (CARS) imaging technique has demonstrated great potential in cancer diagnosis by providing cellular-level resolution images without using exogenous contrast agents. To translate CARS microscopy system into clinical settings, an optical fiber based signal collection scheme must be equipped and an automated image analytics platform must be developed. In this thesis, I first introduce the concept of CARS by showing images acquired from thyroid and parathyroid tissues. I then describe the use of a customized optical fiber bundle to collect and differentiate forward and backward generated CARS signals that contain different structural information. Next, I demonstrate the feasibility of using deep learning algorithm to automatically characterize CARS images. I apply transfer learning on CARS images and achieve 89.2% prediction accuracy in differentiating normal, small-cell carcinoma, adenocarcinoma, and squamous cell carcinoma human lung images. The combination of an optical fiber based microendoscopy and deep learning image classification algorithm will promote CARS imaging towards efficient on-the-spot cancer diagnosis, allowing medical practitioners to obtain essential information in real time and accelerate clinical decision-making. The thesis also shows the generality of the deep learning algorithm developed for image classification in drug discovery. In specific, for automated classification of large volumes of high-content screening images for Alzheimer’s disease; by applying similar transfer learning method on p-Tau images, I separate ineffective, partially-effective, and significantly-effective drugs with high speed and accuracy.