X-Net With Different Loss Functions for Cell Image Segmentation

Haruki Fujii, Hayato Tanaka, Momoko Ikeuchi, Kazuhiro Hotta; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 3793-3800

Convolutional neural network is valid for object segmentation. In recent years, it has been applied to the fields of medicine and cell biology. Each class has a different number of pixels in an image. Therefore, the accuracy of semantic segmentation varies drastically between objects with a large number of pixels and objects with a small number of pixels. In this paper, we propose X-Net that integrates two encoders and decoders to solve this problem. This has the advantage of extracting rich features from two encoders and using two decoders to complement the location information and small objects. By using different loss functions for each decoder, we can use the ensemble of two decoders with different viewpoints. We evaluated our method on the Arabidopsis thaliana cell images and Drosophila cell images. Experimental results show that our method achieved better accuracy than the conventional methods.