Inter- and intra-uncertainty based feature aggregation model for semi-supervised histopathology image segmentation

TL;DR

A novel inter- and intra-uncertainty regularization method to measure and constrain both inter- and intra-inconsistencies in the teacher-student architecture is proposed and a new two-stage network with pseudo-mask guided feature aggregation (PG-FANet) as the segmentation model is proposed.

Abstract

Acquiring pixel-level annotations is often limited in applications such as histology studies that require domain expertise. Various semi-supervised learning approaches have been developed to work with limited ground truth annotations, such as the popular teacher-student models. However, hierarchical prediction uncertainty within the student model (intra-uncertainty) and image prediction uncertainty (inter-uncertainty) have not been fully utilized by existing methods. To address these issues, we first propose a novel inter- and intra-uncertainty regularization method to measure and constrain both inter- and intra-inconsistencies in the teacher-student architecture. We also propose a new two-stage network with pseudo-mask guided feature aggregation (PG-FANet) as the segmentation model. The two-stage structure complements with the uncertainty regularization strategy to avoid introducing extra modules in solving uncertainties and the aggregation mechanisms enable multi-scale and multi-stage feature integration. Comprehensive experimental results over the MoNuSeg and CRAG datasets show that our PG-FANet outperforms other state-of-the-art methods and our semi-supervised learning framework yields competitive performance with a limited amount of labeled data.

Figures (8)

• Figure 1: The overall architecture of the proposed semi-supervised histopathology image segmentation model using two-stage PG-FANet and inter- and intra-uncertainty and consistency regularization. EMA denotes exponential moving average.
• Figure 2: Overview of our (a) PG-FANet with two-stage sub-networks, (b) mask-guided feature enhancement (MGFE) module, (c) multi-scale feature aggregation, and (d) multi-stage feature aggregation. Both stages share the same convolution block (CB). RB$i\_s$ denotes the $i$th residual blocks in stage $s$. RB3$\_s$ and RB4$\_s$ are dilated residual blocks with dilation rates of 2 and 4 respectively to generate feature maps with various receptive fields. Conv denotes the convolutional layer. ASPP denotes the atrous spatial pyramid pooling module. Up denotes the upsampling operation. The size of the output feature maps is given by batch size $\times$ channel size $\times$ height $\times$ width (B $\times$ C $\times$ H $\times$ W).
• Figure 3: Segmentation results on the MoNuSeg dataset with each add-on component in PG-FANet-SSL using 5% labeled data.
• Figure 4: Segmentation results on the MoNuSeg and CRAG datasets using our fully supervised PG-FANet with 100% labeled data and semi-supervised learning with 5%, 10%, 20%, and 50% of the labeled data.
• Figure 5: A representative segmentation outcome achieved with our method on the MoNuSeg and CRAG datasets, compared with results from other state-of-the-art approaches.