Midterm colloquium Shyam Prasadh Sankararaman

21 June 2018 13:30 till 14:00 - Location: lecture room F, 3me - By: DCSC

"Tissue characterization by Deep Learning in Medical Hyperspectral Images"

Hyperspectral imaging (HSI) is an emerging modality in medical imaging applications, that originated in remote sensing environments. It has the capability of acquiring 2-D images across a wide range of wavelengths, constituting a 3-D hypercube of image data and can be a potential tool in medicine, for noninvasive, non-contact disease diagnosis and image-guided surgery. It encodes both spatial and spectral information that can detect subtle changes in the biochemical properties of a tissue, thus revealing the progression of a pathological condition. This study seeks to review some of the research relevant to (tumor) tissue characterization using HSI data. Some of the advantages and disadvantages are also discussed to fully understand how the modality could be put to use.

Deep learning is a neural network based learning method that can be used in medical images to classify, localize or semantically segment an affected tissue, based on training the model on hundreds of annotated images. In the recent years, several new complex models, publicly available databases and an exponential increase in computational power have led to their successful application in medical image analysis. Initial work in this thesis reviews the different approaches for tasks like image classification, tumor detection and segmentation. Furthermore, the most recent research on HSI image analysis (remote sensing and medical diagnosis) using deep learning is summarized and the methods most relevant to this project are specifically highlighted. Subsequent work involves the preparation of training data and labels, identification of suitable deep learning models for hyperspectral image data, deployment of models using TensorFlow and Python, and evaluation of performance of the considered approaches.

Supervisor:
Dr.Ing. R. Van de Plas