Ensemble CNNs-Transformers Fusion Network for Tumor Segmentation in Pathological Images

The advancement of artificial intelligence has rendered deep learning capable of automating the process of diagnosing pathological images with speed, precision, and dependability. However, despite of this progress, existing networks often prioritize enhancing feature extraction in the encoder stage, neglecting the equal importance of the decoder stage’s structural design for recover details. In response to this issue, this study introduces a novel ensemble network architecture that integrates three different up-sampling modules in decoder, which are multi-transposed convolutional sampling, bilinear upsampling, and Swin Transformer upsampling. It also employs Swin Transformer and ConvNeXt as the dual-branch parallel and independent encoders. They can fully capture the local and global dependencies between pathological image features. This network was evaluated on the liver cancer dataset and GLAS dataset, and demonstrated favorable performance indicators against various mainstream networks.