Research On Medical Image Segmentation Based On Feature Encoder-decoder Structure

Medical image segmentation is a significant problem in biomedical image processing.Efficient and automatic segmentation of medical images helps for the diagnosis and treatment of disease,and provides powerful techniques for researchers to better understand the anatomical information of biological tissues.Image noise and intensity non-uniformity are the main barriers for accurate segmentation and robustness of models.In recent years,methods using feature encoder-decoder structures have gradually become the mainstream.However,many methods do not take the characteristics such as multiple modalities and spatial adjacency reslationships in medical images when designing the feature encoder.Moreover,when building an end-toend deep neural network,the design of decoders does not consider the long-term dependency when fusing feature maps form different layers.Therefore,this paper discuss the design of many popular encoder-decoder models and propose two novel methods to further improve the performance and robustness in medical image segmentation.The main work of this paper includes:(1)From the perspective of improving the feature encoder,a feature encoder-decoder method based on multi-modality and adjacency constraint is proposed,namely LSTM-MA.The method considers an image as an undirected graph connected by graph nodes.Combined with multi-modality and adjacency constraint,we design two feature sequence generation ways,i.e.,features with pixel-wise and superpixel-wise adjacency constraint.The LSTM model classifies the generated features into semantic labels to form the segmentation result.The evaluation experiments on BrainWeb and MRBrainS demonstrate that the method is robust to image noise.LSTM-MA with pixel-wise adjacency constraint achieves promising segmentation results,while LSTM-MA with superpixel-wise adjacency constraint reduces the time complexity with a little sacrifice in accuracy.(2)From the perspective of improving the feature decoder,a feature encoder-decoder method based on convolutional RNNs is proposed.We design a novel feature fusion unit called Recurrent Decoding Cell(RDC)which leverages convolutional RNNs to memorize the longterm context information from the previous layers in the decoding phase.An encoder-decoder network,named Convolutional Recurrent Decoding Network(CRDN),is also designed based on RDC for segmenting multi-modality medical images.CRDN adopts CNN backbone to encode image features and decode them hierarchically through a chain of RDCs to obtain the final high-resolution score map.The evaluation experiments on BrainWeb,MRBrainS and HVSMR datasets demonstrate that the introduction of RDC effectively improves the segmentation accuracy as well as reduces the model size,and the proposed CRDN owns its robustness to image noise and intensity non-uniformity in medical images.