Weakly Supervised Breast Ultrasound Image Segmentation Based on Image Selection
Click the Poster to View Full Screen, Right click to save image
Grant: Students Partnering with Faculty (SpF)
Tzu-Han Lin
College:
The Dorothy and George Hennings College of Science, Mathematics, and Technology
Major:
Computer Information Systems
Faculty Research Advisor(s):
Kuan Huang, Daehan Kwak
Abstract:
Automatic segmentation in Breast Ultrasound (BUS)
imaging is vital to BUS computer-aided diagnostic systems. Fully
supervised learning approaches can attain high accuracy, yet they
depend on pixel-level annotations that are challenging to obtain.
As an alternative, weakly supervised learning methods offer a
way to lessen the dependency on extensive annotation require-
ments. Existing weakly supervised learning methods are typically
trained on the entire dataset, but not all samples are effective
in training a robust image segmentation model. To overcome
this challenge, a new weakly supervised learning approach for
BUS image segmentation has been developed. Our framework
includes three key contributions: 1) A novel image selection
method using Class Activation Maps is employed to identify
high-quality candidates for generating pseudo labels; 2) The
‘Segment Anything’ is utilized for pseudo-label generation; 3)
A segmentation model is trained using a Mean Teacher method,
incorporating both pseudo-labeled and non-labeled images. The
proposed framework is evaluated on a public BUS image dataset
and achieves an Intersection over Union score that is 82.9% of
what is attained by fully supervised methods.