Pathological Image Feature Representation Based On Few-Shot Learning

With the continuous development of medical imaging technology and artificial intelligence algorithms,computer-aided diagnosis has attracted more and more attention.As one of the key technologies,medical image segmentation has gradually become a research hotspot in academia.However,medical images usually contain rich information of tissue structure,and there is a great difference in data distribution between medical images and natural images,leading to a poor generalization of segmentation model.The lack of data annotation further limits the performance of the model.The In this paper,three aspects are explored,including domain adaptation,structured feature representation and few-shot learning in medical image segmen-tation.The main contributions can be summarized as follows:1.For the domain adaptation in pathological image segmentation,a domain adaptation algorithm combining attention module and generative adversarial network structure is proposed.By capturing high-dimensional features with domain invariance,the generalization of the model in cross-domain tasks is improved.2.For pathological image segmentation based on structured feature representation,a multi-resolution information fusion module and weighted stitching post-processing method are proposed,which can effectively solve the block artifact problem caused by patch stitching.Moreover,the structured feature extraction from pathological images is achieved by capturing the spatial correlation of multi-resolution images.3.In the medical image segmentation based on few shot learning,a training strategy based on self-supervised learning is proposed.By introducing the local prototype representation of images,the fine-grained features in medical images are explicitly modeled,which improves the robustness of the image segmentation in case of data starvation.The effectiveness of the proposed algorithm is verified by ablation experiments and comparision with state-of-the-art algorithms on multiple public datasets.