289 LG: Dr. Javier Villafruela from University of Calgary will be presenting "Practical Biomedical Texture Classification with Deep Learning"

Friday, October 7, 2016 - 10:00
Javier Villafruela
During the last years, several studies have shown that texture analysis from medical images can characterize tumor heterogeneity and predict clinical outcomes for several diseases based on local variations of intensities. Texture classification involves four steps: lesion segmentation, feature extraction, feature selection and learning. Lesion segmentation is considered to be the bottleneck step, since manually delineation is a high time-consuming task. During the last years, advanced multi-atlas-based segmentation propagation frameworks have been demonstrated to be a robust and accurate alternative to manual contouring. A wide variety of algorithms for constructing statistical image signatures have been proposed, such as gray-level co-occurrence matrix, local binary patterns, wavelet transform, etc. In order to improve the accuracy of the predictions, only the most informative and non-redundant features are selected to train a classifier. Recently, deep learning has gained a lot of momentum and interest from the pattern recognition community. The most recent radiomics studies from the Medical Image Processing Laboratory of the University of Calgary will be presented in this seminar, including findings on clinical outcome prediction for sarcomas and non-small cell lung cancer, classification of cardioembolic and arteriosclerotic strokes using apparent diffusion coefficient mapping, and detection of early ischemic changes on non-contrast CT scans.