Semantic knowledge distillation for conjunctival goblet cell segmentation

Recently, as the usage of electronic devices increase, modern people suffer from eye diseases. We analyzed goblet cells of wide-field fluorescence microscopy with a deep learning. In this study, we propose to real-time analysis using knowledge distillation using proposed loss function and optimized network. In the proposed method, residual based U-Net was used as the teacher network to distill knowledge into lightweight E-Net. We train the student network using pixelwise loss and. The proposed method showed 4% improvements in dice-score compared to the lightweight E-Net, and the processing time was decreased to 68% compared to the case where only the teacher network was performed.