An Improved Autoencoder Approach for Nuclei Image Segmentation

Abstract

There is a dire need to enable an early diagnosis system to enhance the therapeutic outcome for patients by applying a medical image analysis application. This study proposes an improved auto-encoder model by integrating Squeeze and Excitation (SE) blocks on the different phases of the model for semantic segmentation, which U-Net inspires. We redesigned the model's skip-connection by utilizing Residual Squeez and Excitation (RSE) by employing SE block in a residual way to reduce the semantic gaps and discrepancy between encoder and decoder features. Then, we integrate the Dense Squeeze and Excitation (DSE) block in the model's bottleneck with a densely connected structure. We increase the model's accuracy compared to vanilla U-Net by integrating the discussed module in the model to enhance its capability for feature extraction and obtain more high-level features from the input feature. To evaluate our model's performance, we conducted our experiment on the 2018 Data Science Bowl dataset and compared it with the different approaches that are inspired by U-Net. Our proposed model achieved the Dice and IoU of 92.15% and 85.92% , respectively, surpassing most of the current stateof-the-art models.