Semi-Supervised Skin Lesion Segmentation with Learning Model Confidence

Abstract

Segmentation of skin lesions is important for disease diagnoses and treatment planning. Over the years, semisupervised methods using pseudo labels have boosted the segmentation performance with limited labeled data and abundant unlabeled data. However, the unreliable targets in pseudo labels might lead to meaningless guidance for unlabeled data. In this paper, to solve this issue, we propose a novel confidence aware semi-supervised learning method based on a mean teacher scheme. Concretely, we design a confidence module to predict the model confidence guided by the True Class Probability. Then in the mean teacher framework, the student model gradually learns trustworthy targets from teacher model. To further improve the segmentation quality, we fine-tune the student model with reliable content in pseudo labels. We conduct extensive experiments on 2018 ISIC skin lesion segmentation dataset and our method outperforms other state-of-the-art semi-supervised approaches.

Confidence-aware cross-supervised model for semi-supervised skin lesion segmentation

“…By controlling for all other variables in Eqs. (14), we compare the JA index with different λ from 1 to 8. As shown in Table 3, setting the trade-off weight λ to 3 gives the best results on the ISIC2017 dataset.…”

Learning with Limited Annotations: A Survey on Deep Semi-Supervised Learning for Medical Image Segmentation

“…Yu et al [52] extend the mean teacher paradigm with an uncertainty estimation strategy through Monte Carlo dropout [53]. Xie et al [54] add a confidence-aware module to learn the model confidence under the guidance of labeled data. Luo et al [55], [56] calculate uncertainty using pyramid predictions in one forward pass and proposed an multi-level uncertainty rectified pyramid consistency regularization.…”

Deep semi-supervised learning via dynamic anchor graph embedding in latent space

“…Finally, we apply (3) to integrate the teacher model with more diverse "students" generated by (9). Notice that we apply ( 9) and ( 3) to the whole parameter set θ, which could increases the quality of consistency target f (x, ν ) in (5).…”

“…Deep neural networks leveraging a large number of labeled data have achieved remarkable performance on computer vision [1,2] and natural language pro-cessing [3,4] applications in recent years. However, labeling numerous data manually is extremely expensive or even impossible for many practical applications (e.g., medical image segmentation [5], destructive product testing [6]). In these situations, deep learning models are prone to overfitting and generalize badly on new data, due to that the labeled training data are scarce.…”