PROJECT TITLE :

On Smart Gaze based Annotation of Histopathology Images for Training of Deep Convolutional Neural Networks

ABSTRACT:

To fully realize the potential of Deep Learning in histopathology applications, a bottleneck that needs to be overcome is the lack of availability of large training datasets. Even though the digitization of slides through the use of whole slide imaging scanners has made the process of data acquisition much quicker, the labeling of virtual slides still requires a significant amount of time on the part of pathologists. There is a possibility that using eye gaze annotations will speed up the process of slide labeling. This study investigates whether or not eye gaze labeling is a viable alternative to traditional manual labeling for the purpose of training object detectors, as well as timing comparisons between the two methods. In this article, we discuss not only the difficulties that come with gaze-based labeling but also the methods that can be used to refine the coarse data annotations in preparation for subsequent Object Detection. The findings indicate that using gaze tracking for labeling can help save valuable time for pathologists and achieves satisfactory results when utilized in the process of training a deep object detector. We compare the performance gap between deep object detectors trained using hand-labelled data and gaze-labelled data by using the task of localization of Keratin Pearls in cases of oral squamous cell carcinoma as a test case. When compared to 'Bounding-box' based hand-labeling, gaze-labeling required 57.6% less time per label on average, and when compared to 'Freehand' labeling, gaze-labeling required 85% less time per label on average. Both of these figures are averages.