Uncertainty-Aware Self-ensembling Model for Semi-supervised 3D Left Atrium Segmentation

Abstract

Training deep convolutional neural networks usually requires a large amount of labeled data. However, it is expensive and timeconsuming to annotate data for medical image segmentation tasks. In this paper, we present a novel uncertainty-aware semi-supervised framework for left atrium segmentation from 3D MR images. Our framework can effectively leverage the unlabeled data by encouraging consistent predictions of the same input under different perturbations. Concretely, the framework consists of a student model and a teacher model, and the student model learns from the teacher model by minimizing a segmentation loss and a consistency loss with respect to the targets of the teacher model. We design a novel uncertainty-aware scheme to enable the student model to gradually learn from the meaningful and reliable targets by exploiting the uncertainty information. Experiments show that our method achieves high performance gains by incorporating the unlabeled data. Our method outperforms the state-of-the-art semi-supervised methods, demonstrating the potential of our framework for the challenging semi-supervised problems 3 .

A Noise-Robust Framework for Automatic Segmentation of COVID-19 Pneumonia Lesions From CT Images

“…First, to deal with noisy annotations for training CNNs to segment COVID-19 pneumonia lesions, we propose a novel noiserobust Dice loss function, which is a combination and generalization of MAE loss [14] that is robust against noisy labels and Dice loss [18] that is insensitive to foreground-background imbalance. Second, we propose a novel noise-robust learning framework based on self-ensembling of CNNs [20], [21], where an Exponential Moving Average (EMA, a.k.a. teacher) of a model is used to guide a standard model (a.k.a.…”

“…student) to improve the robustness. Differently from previous selfensembling methods for semi-supervised learning [20], [21] and domain adaptation [22], [23], we propose two adaptive mechanisms to better deal with noisy labels: adaptive teacher that suppresses the contribution of the student to EMA when the latter has a large training loss, and adaptive student that learns from the teacher only when the teacher outperforms the student. Thirdly, to better deal with the complex lesions, we propose a novel COVID-19 Pneumonia Lesion segmentation network (COPLE-Net) that uses a combination of max-pooling and average pooling to reduce information loss during downsampling, and employs bridge layers to alleviate the semantic gap between features in the encoder and decoder.…”

“…To further improve the performance of dealing with noisy labels, we integrate our COPLE-Net and L NR-Dice into a selfensembling framework that was originally proposed for semisupervised learning [20], [21]. It consists of a teacher model T and a student model S with the same CNN structure, and T is updated as an Exponential Moving Average (EMA) of S:…”

“…and θ t to θ * t , respectively. Due to EMA, the teacher model T is more stable than the student model S and can be used to supervise S through a consistency loss L c [20], [21]. Let x and y represent a training image and its noisy annotation respectively, the overall loss function for training is:…”

“…In previous self-ensembling methods [20], [21] for semisupervised learning, values of α and λ were manually set as fixed numbers or changed gradually according to the training step t, without considering the performance of S and T and the existence of noisy labels. In our segmentation task, the student model S may have a poor performance at step t due to noisy labels, and updating the teacher model T with a predefined α (e.g., 0.99 in [21]) may lead T to be corrupted by the noisy labels as well.…”