Serial tissue sections are widely used in imaging large tissue volumes. Navigating to each section is indispensable in the automatic imaging process. Nowadays, the locations of sections are labeled manually or semi-manually. Sections are similar and the border is indiscernible if they stick together, which makes it difficult to locate the sections automatically. In this paper, we present frequency-aware instance segmentation framework (FANet), which can extract shape and size information of sections very well. Firstly, FANet uses discrete cosine transform(DCT) . Secondly, each channel extracts an specific frequency component of themselves. Frequency components from all channels is taken as the multi-frequency description of feature map and finally used to model the channel attention. Additionally, we propose a dataset about the serial sections as benchmark, which contains 2708 images in training set and 1193 images in validation set. Experimental results on the benchmark demonstrate our FANet achieves superior performance compared with the current methods. Our code and dataset will be made public.