Keywords: Ultrasound segmentation, Semi-supervised learning, Metric optimization
TL;DR: We propose a method to optimize a non-differentiable anatomical prior which can reduce anatomical errors on labeled and un-labeled data in a semi-supervised manner.
Abstract: Deep convolutional neural networks (CNNs) have had great success for medical imaging segmentation. Many methods attained nearly perfect Dice scores, sometimes within inter-expert variability. However, CNNs require large amounts of labeled data and are not immune to producing anatomically implausible results, especially when applied to ultrasound images. In this paper, we propose a method that tackles both of these problems simultaneously. Our method optimizes anatomical segmentation metrics on both labeled and unlabeled data using a training scheme analogous to adversarial training. Our method allows the optimization of several hand-made non-differentiable metrics for any segmentation model and drastically reduces the number of anatomical errors.
Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: novel methodological ideas without extensive validation
Primary Subject Area: Segmentation
Secondary Subject Area: Learning with Noisy Labels and Limited Data
Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.
Code And Data: The code is available at https://github.com/ThierryJudge/anatomically-constrained-ssl.