B1629
Title: Supervised machine learning for segmentation misclassification in neuroimaging
Authors: Eunchan Bae - University of Pennsylvania (United States) [presenting]
Russell Shinohara - University of Pennsylvania (United States)
Abstract: Recent advancements in machine learning facilitate a deeper understanding of biomedical research. Automatic segmentation in biomedical imaging, the identification of regions of interest, is one of the areas that has flourished with machine learning. Most supervised machine learning algorithms rely on the assumption that gold standard manual labels are correct. However, if the labels or measurements used in model training are inaccurate, supervised algorithms become unreliable. In biomedical imaging, misclassification of labels is common due to inhomogeneous intensities in images, low-resolution images, and manual segmentation variability. Therefore, there is a need to relax the assumptions of no misclassification when building supervised machine learning algorithms. A novel iterative misclassification-adjusting supervised machine learning algorithm (ITEMS) is proposed that estimates the false-positives rates and false-negatives rates of the error-prone labels and simultaneously self-corrects the labels.
