Inception Inspired U-Net for Effective Segmentation of Microscopy Images

Abstract

Biomedical image analysis has vital role in medical diagnosis. Computerized tools for automatic analysis of biomedical images helps the radiologist for the disease identification. This work explores a custom designed deep learning architecture for automatic segmentation of microscopy images. The deep learning architecture uses Inception modules in U-Net type structure for effective segmentation. This newly designed framework is tested on different types of microscopy data to assess segmentation performance. The proposed model's dice similarity index for the DIC-C2DH-HELA and Fluo-C2DL-MSC datasets are 0.9559 and 0.9167 respectively. Further intersection over union for the DIC-C2DH-HELA and Fluo-C2DL-MSC datasets are 0.8783 and 0.74779 respectively.