Segmentation of Liver Cancer Pathology Images using Transformer-CNNs with High-Resolution Semantic Features

Pathological images serve as the ‘gold standard’ for tumor diagnosis. The traditional process of analyzing pathological sections under a microscope is difficult, time-consuming and inefficient. Even expert pathologists may make errors due to subjective judgments. With the advancement of deep learning technology, automatic segmentation of cancerous regions in pathological images using deep learning has greatly reduced the burden on pathologists and the possibility of misdiagnosis or missed diagnosis. The Transformer architecture has proven to be highly successful in capturing global information and modeling long-range dependencies. However, it has limitations in learning local relationships (context). This can result in the loss of local information, such as boundaries and small objects, especially in pathology images with rich details. Additionally, the decoder stage cannot access more high-resolution primary features of pathological images. To address this issue, this study proposes a novel fusion architecture based on Transformer and CNN. It designs a new feature extraction module specifically for high-resolution pathological images, introduces a boundary information extraction module, and integrates a gated attention mechanism. The high-resolution feature extraction module extracts features from the first layer of high-resolution feature maps obtained from the encoder, effectively expanding the network’s receptive field and supplementing lost local information. The boundary information extraction module aggregates boundary features from semantic information of both low-level and high-level features to better preserve boundary details and recalibrate the localization of segmented objects. The network’s generalization performance is evaluated on the LiTs17 liver segmentation dataset and compared with various mainstream networks, achieving good performance metrics. Furthermore, the network is compared with similar networks on a liver cancer pathology dataset provided by Eastern Hepatobiliary Hospital, confirming the effectiveness of the proposed network.