SpiralNet: Two-stage recursive-CNN for microscopy image segmentation

Abstract

Microscopy image segmentation demands a higher precision level than segmentation for natural images. Meticulous accuracy is required for medical applications. SpiralNet is designed as a new segmentation method allowing to segment microscopy images of complex shapes with high attention to details simulating human perception. The method is able to perform both instance and semantic segmentation. SpiralNet consists of two stages, the first stage crops the initial image into smaller regions and with a scoring network filters out regions without objects. The second stage takes each region and fully segments it with a recursive segmentation network. Afterwards, the predicted regions are merged into the final full prediction mask. SpiralNet outperforms U-Net with a 0:969 F1 score versus U-Net 0:965 on the test subset, segmenting more accurate individual object shapes and showing better separation between connected objects. Even though SpiralNet showed great instance and semantic segmentation performance, there are still various ways to improve the method. For instance, with parallel segmentation of several regions, adding attention or changing the number of skip modules. Additionally, future work will study the application of SpiralNet to other datasets.