Publication

Abstract : Encoder-decoder architectures have been extensively used for semantic segmentation tasks, including biomedical image segmentation. Most architectures utilize UNet as the base deep learning encoder-decoder network for biomedical image segmentation. To extract global and local features, multi-scale feature extraction using convolution involving different kernel sizes is very important for an effective analysis. UNet3+ can extract multi-level features from different layers of the encoder-decoder, which could be appropriately concatenated with decoder blocks. This paper leverages the combination of multi-scale features from convolution to extract higher-level semantics and multi-level hierarchical features from different blocks of the encoder-decoder as performed with UNet3+. We also incorporate multi-head attention in the bottleneck to further enhance extracted features from convolution, thus making a UNet architecture combining Convolution and Attention for Segmentation (CASe_UNet), combined with a hybrid loss for reducing precision and recall in pixel classification. The proposed architecture also tries to encounter class imbalance by assigning class weights inversely proportional to the probability distribution of each class. The approach is trained and validated on LiTS 2017 to identify tumors in the Liver and EndoVis 2018 Medical Instrument Segmentation data. Experimental results, along with ablation experiments, show that the proposed architecture achieves good results with an IOU and Dice coefficient of 0.9438 and 0.8684, respectively, on the EndoVis 2018 dataset and the benchmark IoU and Dice coefficient of 0.9166 and 0.9482 on LiTS 2017 data.