Abstract:
Unavailability of large training datasets is a bottleneck that needs to be
overcome to realise the true potential of deep learning in histopathology
applications. Although slide digitisation via whole slide imaging scanners has
increased the speed of data acquisition, labelling of virtual slides requires a
substantial time investment from pathologists. This makes generating large,
correctly-labelled slide datasets an expensive, time-consuming and laborious
exercise. Eye gaze annotations have the potential to speed up the slide
labelling process. This work explores the viability of using eye gaze labelling
as compared to conventional hand based labelling techniques for training
object detectors. A low-cost gaze tracking device is used to track the gaze of
a pathologist working with virtual slides on a computer screen. Challenges
associated with gaze based labelling and techniques to refine the coarse data
for subsequent object detection are also discussed. Results demonstrate that
gaze tracking based labelling can save valuable time of the pathologist while
performance of deep learning algorithms trained to detect Keratin Pearls in
oral cancer Whole Slide Images (WSI) using gaze annotations is comparable
to deep learning models obtained from hand annotations.
