An Improved U-Net Model for Simultaneous Nuclei Segmentation and Classification

Nuclei segmentation and classification of pathological images are the first steps in automatic cancer prognosis and grading. Recent methods use the U-Net variant models to segment and classify nuclei. However, the U-Net variant models suffer from frequency domain information loss when extracting nuclear features due to frequency bias and position shift. In this paper, we propose an improved U-Net model, which mainly includes two plug-and-play modules to tackle the above problem. First, the Frequency-aware Feature Enhancement (FFE) module adopts frequency domain learning to enhance the features of U-Net learned in the image domain. Specifically, the FFE module obtains a learnable calibration coefficient matrix by supervising the spectrum of the received feature. The spectrum is then enhanced by the matrix and converted back to the image domain to alleviate frequency bias. Second, the Position-aware Feature Aggregation (PFA) module is used to mitigate information loss. The PFA module aligns each of the multiscale features with rich semantic information and the top-layer feature with rich position information. The aligned features are then aggregated to provide a strong information flow for prediction. Experimental results on the PanNuke dataset demonstrate that our model effectively improves the nuclei segmentation and classification performance of the U-Net variant models. Compared with Hover-Net, our method improves mPQ by 3.95% and bPQ by 1.12%.