Semi-supervised Deep Learning for Fully Convolutional Networks

Abstract

Abstract. Deep learning usually requires large amounts of labeled training data, but annotating data is costly and tedious. The framework of semi-supervised learning provides the means to use both labeled data and arbitrary amounts of unlabeled data for training. Recently, semisupervised deep learning has been intensively studied for standard CNN architectures. However, Fully Convolutional Networks (FCNs) set the state-of-the-art for many image segmentation tasks. To the best of our knowledge, there is no existing semi-supervised learning method for such FCNs yet. We lift the concept of auxiliary manifold embedding for semisupervised learning to FCNs with the help of Random Feature Embedding. In our experiments on the challenging task of MS Lesion Segmentation, we leverage the proposed framework for the purpose of domain adaptation and report substantial improvements over the baseline model.

Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis

“…Another popular strategy is to use the unlabeled data to better estimate the distribution of the data, and as such, regularize the classifier. Graph-based methods and semi-supervised SVMs fall under this Reference Application SSL category Brain Song et al (2009) tumor segmentation graph-based Iglesias et al (2010) skull stripping self-training Filipovych et al (2011) classification of MCI semi-supervised SVM Batmanghelich et al (2011) classification of AD, MCI graph-based Xie et al (2013) tissue segmentation graph-based Meier et al (2014) tumor segmentation graph-based Dittrich et al (2014) fetal brain segmentation self-training Wang et al (2014) lesion segmentation self-training, active AD classification graph-based Baur et al (2017) MS lesion segmentation graph-based Moradi et al (2015) classification of MCI semi-supervised SVM Eye Adal et al (2014) microaneurysm detection self-training Mahapatra (2016) optic disc missing annotation prediction self-training, graph-based Breast Sun et al (2016) mass classification co-training Heart Zuluaga et al (2011) detection of vascular lesions self-training Bai et al (2017) cardiac segmentation self-training Wang et al (2017) aneurysm volume estimation graph-based Lung Prasad et al (2009) segmentation of emphysema in CT self-training / co-training, active van Rikxoort et al (2010) classification of tuberculosis patterns in CT self-training / co-training Abdomen Tiwari et al (2010) classification of cancerous areas in prostate graph-based Park et al (2014) prostate segmentation graph-based, active Borga et al (2016) liver segmentation graph-based Mahapatra (2016) predicting missing expert annotations of Crohn's disease self-training, graph-based Histology and microscopy Singh et al (2011) cell type classification in microscopy self-training Parag et al (2014) cell type segmentation in microscopy graph-based, active Xu et al (2016) neuron segmentation in microscopy graph-based Su et al (2016) cell segmentation in microscopy graph-based, active Multiple Gass et al (2012) segmentation in two applications graph-based…”

Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation

“…The semi-supervised framework suggested by Baur et al (2017) consist of a U-Net with two loss functions: a Dice-based segmentation loss that is computed based on the labeled data; and an embedding loss, which, given a batch of labeled and unlabeled data, brings the feature embedding of Table 6: Semi-supervised methods for medical image segmentation. The suggested methods combine the segmentation task with an unsupervised task, allowing the model to use both labeled and unlabeled images during training.…”

“…Unsupervised task Bai et al (2017) Embedding consistency Zhang et al (2017b) Image classification Sedai et al (2017) Image reconstruction Baur et al (2017) Manifold learning Chartsias et al (2018) Image reconstruction Huo et al (2018a) Image synthesis Zhao et al (2019) Image registration Li et al (2019) Transformation consistency the same-class pixels as close as possible while pushing apart the feature embedding of the pixels from different classes. To identify same-class pixels between labeled and unlabeled images, the authors assume the availability of a noisy label prior for unlabeled images.…”

Deep Semi-supervised Segmentation with Weight-Averaged Consistency Targets

“…Only a few works were developed in the context of semi-supervised segmentation, especially in the field of medical imaging. Only recently, a U-Net was used as auxiliary embedding in [19], however, with focus on domain adaptation and using a private dataset.…”