Multi-feature Similarity Based Deep Learning Framework for Semantic Segmentation

Liver tumor is one of the significant causes of death among men and women, but it is confirmed that early detection of the disease ensures the long survival of the patient. In our research, a hybrid of Multi-feature pyramid based U-Net, short skip connections and a Feature similarity module are proposed for early tumor detection. The proposed algorithm focuses on improving the tumor segmentation performance with fewer training parameters. The robustness of the proposed algorithm is claimed on the basis of the dice score coefficient of tumor segmentation. We have achieved a dice score of 0.753 and 0.950 on tumor and liver, respectively on the Liver Tumor Segmentation (LiTS) dataset. In comparison with earlier models, our model has achieved a higher dice coefficient with less training time with nearly 6 million learnable parameters.