Many important information in medical research and clinical diagnosis are obtained from medical images. Among them, digital pathology images can provide detailed tissue structure and cellular information, which has become the gold standard for clinical tumor diagnosis. With the development of neural networks, computer-aided diagnosis presents the identification results of various cell nuclei to doctors, which facilitates the identification of cancerous regions. However, deep learning models require a large amount of annotated data. Pathology images are expensive and difficult to obtain, and insufficient annotation data can easily lead to biased results. In addition, when current models are evaluated on an unknown target domain, there are errors in the predicted boundaries. Based on this, this study proposes a feature alignment-based detail recognition strategy for pathology image segmentation (FASNet). It consists of a preprocessing model and a segmentation network (UNW). The UNW network performs instance normalization and categorical whitening of feature images by inserting semantics-aware normalization and semantics-aware whitening modules into the encoder and decoder, which achieves the compactness of features of the same class and the separation of features of different classes. The FASNet method can identify the feature detail information more efficiently, and thus differentiate between different classes of tissues effectively. The experimental results show that the FASNet method has a Dice Similarity Coefficient (DSC) value of 0.844. It achieves good performance even when faced with test data that does not match the distribution of the training data. Code: https://github.com/zlf010928/FASNet.git.
Keywords: Assisted diagnosis; Deep learning; Digital pathology images; Feature alignment; Insufficient annotation sets.
© 2024 The Authors. Published by Elsevier Ltd.