Summary: In Proceedings, 2002 Medical Image Computing and Computer Assisted
Intervention (MICCAI), Montreal, 2003.
Weakly-Supervised Segmentation of
Non-Gaussian Images via Histogram Adaptation
Jonas August (firstname.lastname@example.org) and Takeo Kanade (email@example.com)
Robotics Institute, Carnegie Mellon University, Pittsburgh, PA
Abstract. Here we apply an active contour model that allows for arbi-
trary intensity distributions inside and outside the boundary of an object
to be segmented in an image. Computationally, we estimate intensity his-
tograms both inside and outside the current boundary estimate, and use
these histograms to define an image energy as their log-likelihood ratio.
Training the model with accurate example segmentations is unnecessary;
initialization with a crude, user-provided segmentation is sufficient.
Speckle and other forms of noise common in medical images is non-Gaussian,
and yet active contour and level-set techniques applied to segment these images
have assumed Gaussian noise . To overcome this limitation, we allow for ar-
bitrary intensity distributions inside and outside the boundary of an object to
A key barrier to applying non-Gaussian statistical models is parameter esti-
mation. We take a weakly-supervised approach in which region statistics from