Unsupervised Adaptive Lesion Segmentation with Similarity Based Cross Pseudo Supervision

Notable advancement has been made in lesion segmentation based on deep learning. Nonetheless, the effectiveness of deep learning heavily relies on the availability of a substantial amount of high-quality labeled data, which is labor-intensive and expensive to require. Moreover, when applied to a different domain, deep learning-based lesion segmentation methods often experience a decline in performance due to variations in device vendors and patient populations. The unsupervised domain adaptation method is proposed to tackle the domain discrepancy problem. There are certain similarities between semi-supervised learning and unsupervised domain adaptation. From an application perspective, semi-supervised learning models can be used to solve unsupervised domain adaptation tasks.To this end, we introduce the cross-pseudo-supervision mechanism in the semi-supervised field for unsupervised adaptive lesion segmentation. However, the method ignores the similar characteristics in medical images. To fully leverage the semantic similarity between domains, a novel similarity-based cross-pseudo-supervised lesion segmentation model is proposed. The proposed model employs a U-Net encoder, trained on source domain, to extract features from both the source and target domains. Subsequently, cosine similarity is computed based on these extracted features. The obtained similarity scores are then employed to weigh the source samples during the training process. The effectiveness of the proposed method is demonstrated through cross-domain experiments conducted on two publicly available COVID-19 lesion segmentation datasets.