Pyramidal Nonlocal Network for Histopathological Image of Breast Lymph Node Segmentation

Abstract

The convolutional neural networks (CNNs) are frequently used in the segmentation of histopathological whole slide image- (WSI) acquired breast lymph nodes. The first layers in deep network architectures generally encode the geometric and color properties of objects in the training set, while the last layers encode the distinctive and detailed properties between classes. Modern segmentation approaches (DeepLabV3+, SegNet, PSPNet) are realized by evaluating these layers together. However, having a high parameter space of all these networks increases the calculation costs and prevents the researchers from working more effectively. In this study, we present a new pyramid-structured segmentation network (NonLocalSeg). Although the proposed network has low parameter space, its segmentation performance is similar to current architectures. The integration of the Nonlocal Module (NLM-a form of attention mechanism) or Asymmetric Pyramid Nonlocal block (APNB) into classical pyramid-built architectures has led to the reduction of network depth and narrowing of the parameter space while enabling coding of low and high image features. These mechanisms suppressed the unfocused background image, emphasizing the focused foreground object. As a result of a series of ablation experiments carried out, it is seen that the NLM and APNL mechanisms give the succeeded results. Although the network architectures adapting these mechanisms contain fewer parameters than current networks, it is observed that they have a similar accuracy (mean intersection over union [IoU]) range.

Keywords

• Deep learning
• Histopathological image segmentation
• Nonlocal network
• Machine learning