A Novel Unet Decoding Strategy for Cervical Cell Mass Segmentation

A novel Unet decoding strategy for segmentation of cervical cell mass is proposed, which is inspired by the sub-pixel convolution. Our proposed method avoids the artificial information added by interpolation method and the noise introduced by transpose convolution method due to padding. After the implementation of the proposed method on our in-house dataset, the quantitative evaluation shows that the sub-pixel convolution has not only higher metric score but also the stable performance, and our proposed method can be used as a promising decoding method for Unet to segment the cellular clump of the cervical cell images