ConDANet: Contourlet Driven Attention Network for Automatic Nuclei Segmentation in Histopathology Images

Abstract

Cell nuclei segmentation, the task of identifying the precise boundary of the nucleus in each cell in a histopathology image, is a rudimentary task prior to single-cell analysis. While addressing this task, two of the major challenges are the precise delineation of the small-shaped nucleus structure and the characterization of the edge region of the nucleus for both the regular and deformed nucleus. To overcome these challenges, a contourlet driven attention network, namely ConDANet, is developed which incorporates a novel attention mechanism along with a content-preserving sampling strategy. Instead of using conventional spatial or channel-wise attention, a controlling signal driven attention mechanism is proposed, which is capable of extracting fine edge details of the nuclei regions by focusing on the relevant edge or boundary information of the nuclei structures. The contourlet transform based controlling signal generation scheme is proposed, which not only exploits the advantage of the multi-scale time frequency localization properties of wavelets but also provides a high degree of directionality. Additionally, the wavelet pooling strategy is incorporated to the network which preserves the textural content of the nucleus in histopathology images and prevents the loss of information in the subsequent sampling operation of the encoder-decoder part of the network. Finally, the proposed method is employed for analyzing three publicly available histopathology datasets to manifest its effectiveness in segmenting nuclei from cellular images extracted from a wide variety of organs and patients. Dice scores obtained in these three datasets are 88.9%, 81.71%, and 75.12% which are found to be superior compared to some state-of-the-art methods.

Keywords

• Nucleus segmentation
• attention mechanism
• contourlet transform
• wavelet pooling
• encoder-decoder network