US 11,854,208 B2
Systems and methods for trainable deep active contours for image segmentation
Demetri Terzopoulos, Los Angeles, CA (US); and Ali Hatamizadeh, Los Angeles, CA (US)
Assigned to The Regents of the University of California, Oakland, CA (US)
Filed by The Regents of the University of California, Oakland, CA (US)
Filed on Jan. 14, 2021, as Appl. No. 17/149,432.
Claims priority of provisional application 62/961,579, filed on Jan. 15, 2020.
Prior Publication US 2021/0217178 A1, Jul. 15, 2021
Int. Cl. G06T 7/149 (2017.01); G06T 7/11 (2017.01); G06N 3/04 (2023.01); G06N 3/08 (2023.01)
CPC G06T 7/149 (2017.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01); G06T 7/11 (2017.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/20116 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method for generating image segmentations from an input image, the method comprising:
receiving an input image;
providing the input image to a Convolutional Neural Network (CNN) backbone;
generating, using the CNN backbone, a set of parameter maps λ_1 and λ_2 from the input image;
generating, using the CNN backbone, an initialization map from the input image;
receiving, using an automatically differentiable Active Contour Model (ACM) comprising L layers, the parameter maps λ_1 and λ_2 and the initialization map;
generating, using the differentiable ACM, an image segmentation based on the set of parameter maps λ_1 and λ_2 and the initialization map;
comparing the image segmentation with a ground-truth label of the input image to compute a set of one or more losses; and
backpropagating the set of one or more losses to update the differentiable ACM and the CNN backbone.